VirtualBox

source: vbox/trunk/include/iprt/asm-amd64-x86.h

Last change on this file was 99739, checked in by vboxsync, 12 months ago

*: doxygen corrections (mostly about removing @returns from functions returning void).
  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 78.5 KB
Line 
1/** @file
2 * IPRT - AMD64 and x86 Specific Assembly Functions.
3 */
4
5/*
6 * Copyright (C) 2006-2023 Oracle and/or its affiliates.
7 *
8 * This file is part of VirtualBox base platform packages, as
9 * available from https://www.virtualbox.org.
10 *
11 * This program is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU General Public License
13 * as published by the Free Software Foundation, in version 3 of the
14 * License.
15 *
16 * This program is distributed in the hope that it will be useful, but
17 * WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * General Public License for more details.
20 *
21 * You should have received a copy of the GNU General Public License
22 * along with this program; if not, see <https://www.gnu.org/licenses>.
23 *
24 * The contents of this file may alternatively be used under the terms
25 * of the Common Development and Distribution License Version 1.0
26 * (CDDL), a copy of it is provided in the "COPYING.CDDL" file included
27 * in the VirtualBox distribution, in which case the provisions of the
28 * CDDL are applicable instead of those of the GPL.
29 *
30 * You may elect to license modified versions of this file under the
31 * terms and conditions of either the GPL or the CDDL or both.
32 *
33 * SPDX-License-Identifier: GPL-3.0-only OR CDDL-1.0
34 */
35
36#ifndef IPRT_INCLUDED_asm_amd64_x86_h
37#define IPRT_INCLUDED_asm_amd64_x86_h
38#ifndef RT_WITHOUT_PRAGMA_ONCE
39# pragma once
40#endif
41
42#include <iprt/types.h>
43#include <iprt/assert.h>
44#include <iprt/x86-helpers.h>
45#if !defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86)
46# error "Not on AMD64 or x86"
47#endif
48
49#if defined(_MSC_VER) && RT_INLINE_ASM_USES_INTRIN
50/* Emit the intrinsics at all optimization levels. */
51# include <iprt/sanitized/intrin.h>
52# pragma intrinsic(_ReadWriteBarrier)
53# pragma intrinsic(__cpuid)
54# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2010 /*?*/
55# pragma intrinsic(__cpuidex)
56# endif
57# pragma intrinsic(_enable)
58# pragma intrinsic(_disable)
59# pragma intrinsic(__rdtsc)
60# pragma intrinsic(__readmsr)
61# pragma intrinsic(__writemsr)
62# pragma intrinsic(__outbyte)
63# pragma intrinsic(__outbytestring)
64# pragma intrinsic(__outword)
65# pragma intrinsic(__outwordstring)
66# pragma intrinsic(__outdword)
67# pragma intrinsic(__outdwordstring)
68# pragma intrinsic(__inbyte)
69# pragma intrinsic(__inbytestring)
70# pragma intrinsic(__inword)
71# pragma intrinsic(__inwordstring)
72# pragma intrinsic(__indword)
73# pragma intrinsic(__indwordstring)
74# pragma intrinsic(__invlpg)
75# pragma intrinsic(__wbinvd)
76# pragma intrinsic(__readcr0)
77# pragma intrinsic(__readcr2)
78# pragma intrinsic(__readcr3)
79# pragma intrinsic(__readcr4)
80# pragma intrinsic(__writecr0)
81# pragma intrinsic(__writecr3)
82# pragma intrinsic(__writecr4)
83# pragma intrinsic(__readdr)
84# pragma intrinsic(__writedr)
85# ifdef RT_ARCH_AMD64
86# pragma intrinsic(__readcr8)
87# pragma intrinsic(__writecr8)
88# endif
89# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2005
90# pragma intrinsic(__halt)
91# endif
92# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2008
93/*# pragma intrinsic(__readeflags) - buggy intrinsics in VC++ 2010, reordering/optimizers issues
94# pragma intrinsic(__writeeflags) */
95# pragma intrinsic(__rdtscp)
96# endif
97# if defined(RT_ARCH_AMD64) && RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2015 /*?*/
98# pragma intrinsic(_readfsbase_u64)
99# pragma intrinsic(_readgsbase_u64)
100# pragma intrinsic(_writefsbase_u64)
101# pragma intrinsic(_writegsbase_u64)
102# endif
103# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2013
104# pragma intrinsic(__lidt)
105# pragma intrinsic(__sidt)
106# pragma intrinsic(_lgdt)
107# pragma intrinsic(_sgdt)
108# endif
109#endif
110
111
112/*
113 * Undefine all symbols we have Watcom C/C++ #pragma aux'es for.
114 */
115#if defined(__WATCOMC__) && ARCH_BITS == 16
116# include "asm-amd64-x86-watcom-16.h"
117#elif defined(__WATCOMC__) && ARCH_BITS == 32
118# include "asm-amd64-x86-watcom-32.h"
119#endif
120
121
122/** @defgroup grp_rt_asm_amd64_x86 AMD64 and x86 Specific ASM Routines
123 * @ingroup grp_rt_asm
124 * @{
125 */
126
127/** @todo find a more proper place for these structures? */
128
129#pragma pack(1)
130/** IDTR */
131typedef struct RTIDTR
132{
133 /** Size of the IDT. */
134 uint16_t cbIdt;
135 /** Address of the IDT. */
136#if ARCH_BITS != 64
137 uint32_t pIdt;
138#else
139 uint64_t pIdt;
140#endif
141} RTIDTR, RT_FAR *PRTIDTR;
142#pragma pack()
143
144#pragma pack(1)
145/** @internal */
146typedef struct RTIDTRALIGNEDINT
147{
148 /** Alignment padding. */
149 uint16_t au16Padding[ARCH_BITS == 64 ? 3 : 1];
150 /** The IDTR structure. */
151 RTIDTR Idtr;
152} RTIDTRALIGNEDINT;
153#pragma pack()
154
155/** Wrapped RTIDTR for preventing misalignment exceptions. */
156typedef union RTIDTRALIGNED
157{
158 /** Try make sure this structure has optimal alignment. */
159 uint64_t auAlignmentHack[ARCH_BITS == 64 ? 2 : 1];
160 /** Aligned structure. */
161 RTIDTRALIGNEDINT s;
162} RTIDTRALIGNED;
163AssertCompileSize(RTIDTRALIGNED, ((ARCH_BITS == 64) + 1) * 8);
164/** Pointer to a an RTIDTR alignment wrapper. */
165typedef RTIDTRALIGNED RT_FAR *PRIDTRALIGNED;
166
167
168#pragma pack(1)
169/** GDTR */
170typedef struct RTGDTR
171{
172 /** Size of the GDT. */
173 uint16_t cbGdt;
174 /** Address of the GDT. */
175#if ARCH_BITS != 64
176 uint32_t pGdt;
177#else
178 uint64_t pGdt;
179#endif
180} RTGDTR, RT_FAR *PRTGDTR;
181#pragma pack()
182
183#pragma pack(1)
184/** @internal */
185typedef struct RTGDTRALIGNEDINT
186{
187 /** Alignment padding. */
188 uint16_t au16Padding[ARCH_BITS == 64 ? 3 : 1];
189 /** The GDTR structure. */
190 RTGDTR Gdtr;
191} RTGDTRALIGNEDINT;
192#pragma pack()
193
194/** Wrapped RTGDTR for preventing misalignment exceptions. */
195typedef union RTGDTRALIGNED
196{
197 /** Try make sure this structure has optimal alignment. */
198 uint64_t auAlignmentHack[ARCH_BITS == 64 ? 2 : 1];
199 /** Aligned structure. */
200 RTGDTRALIGNEDINT s;
201} RTGDTRALIGNED;
202AssertCompileSize(RTIDTRALIGNED, ((ARCH_BITS == 64) + 1) * 8);
203/** Pointer to a an RTGDTR alignment wrapper. */
204typedef RTGDTRALIGNED RT_FAR *PRGDTRALIGNED;
205
206
207/**
208 * Gets the content of the IDTR CPU register.
209 * @param pIdtr Where to store the IDTR contents.
210 */
211#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2013
212RT_ASM_DECL_PRAGMA_WATCOM(void) ASMGetIDTR(PRTIDTR pIdtr);
213#else
214DECLINLINE(void) ASMGetIDTR(PRTIDTR pIdtr)
215{
216# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2013
217 __sidt(pIdtr);
218# elif RT_INLINE_ASM_GNU_STYLE
219 __asm__ __volatile__("sidt %0" : "=m" (*pIdtr));
220# else
221 __asm
222 {
223# ifdef RT_ARCH_AMD64
224 mov rax, [pIdtr]
225 sidt [rax]
226# else
227 mov eax, [pIdtr]
228 sidt [eax]
229# endif
230 }
231# endif
232}
233#endif
234
235
236/**
237 * Gets the content of the IDTR.LIMIT CPU register.
238 * @returns IDTR limit.
239 */
240#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2013
241RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMGetIdtrLimit(void);
242#else
243DECLINLINE(uint16_t) ASMGetIdtrLimit(void)
244{
245 RTIDTRALIGNED TmpIdtr;
246# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2013
247 __sidt(&TmpIdtr);
248# elif RT_INLINE_ASM_GNU_STYLE
249 __asm__ __volatile__("sidt %0" : "=m" (TmpIdtr.s.Idtr));
250# else
251 __asm
252 {
253 sidt [TmpIdtr.s.Idtr]
254 }
255# endif
256 return TmpIdtr.s.Idtr.cbIdt;
257}
258#endif
259
260
261/**
262 * Sets the content of the IDTR CPU register.
263 * @param pIdtr Where to load the IDTR contents from
264 */
265#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2013
266RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetIDTR(const RTIDTR RT_FAR *pIdtr);
267#else
268DECLINLINE(void) ASMSetIDTR(const RTIDTR RT_FAR *pIdtr)
269{
270# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2013
271 __lidt((void *)pIdtr);
272# elif RT_INLINE_ASM_GNU_STYLE
273 __asm__ __volatile__("lidt %0" : : "m" (*pIdtr));
274# else
275 __asm
276 {
277# ifdef RT_ARCH_AMD64
278 mov rax, [pIdtr]
279 lidt [rax]
280# else
281 mov eax, [pIdtr]
282 lidt [eax]
283# endif
284 }
285# endif
286}
287#endif
288
289
290/**
291 * Gets the content of the GDTR CPU register.
292 * @param pGdtr Where to store the GDTR contents.
293 */
294#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2013
295RT_ASM_DECL_PRAGMA_WATCOM(void) ASMGetGDTR(PRTGDTR pGdtr);
296#else
297DECLINLINE(void) ASMGetGDTR(PRTGDTR pGdtr)
298{
299# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2013
300 _sgdt(pGdtr);
301# elif RT_INLINE_ASM_GNU_STYLE
302 __asm__ __volatile__("sgdt %0" : "=m" (*pGdtr));
303# else
304 __asm
305 {
306# ifdef RT_ARCH_AMD64
307 mov rax, [pGdtr]
308 sgdt [rax]
309# else
310 mov eax, [pGdtr]
311 sgdt [eax]
312# endif
313 }
314# endif
315}
316#endif
317
318
319/**
320 * Sets the content of the GDTR CPU register.
321 * @param pGdtr Where to load the GDTR contents from
322 */
323#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2013
324RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetGDTR(const RTGDTR RT_FAR *pGdtr);
325#else
326DECLINLINE(void) ASMSetGDTR(const RTGDTR RT_FAR *pGdtr)
327{
328# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2013
329 _lgdt((void *)pGdtr);
330# elif RT_INLINE_ASM_GNU_STYLE
331 __asm__ __volatile__("lgdt %0" : : "m" (*pGdtr));
332# else
333 __asm
334 {
335# ifdef RT_ARCH_AMD64
336 mov rax, [pGdtr]
337 lgdt [rax]
338# else
339 mov eax, [pGdtr]
340 lgdt [eax]
341# endif
342 }
343# endif
344}
345#endif
346
347
348
349/**
350 * Get the cs register.
351 * @returns cs.
352 */
353#if RT_INLINE_ASM_EXTERNAL
354RT_ASM_DECL_PRAGMA_WATCOM(RTSEL) ASMGetCS(void);
355#else
356DECLINLINE(RTSEL) ASMGetCS(void)
357{
358 RTSEL SelCS;
359# if RT_INLINE_ASM_GNU_STYLE
360 __asm__ __volatile__("movw %%cs, %0\n\t" : "=r" (SelCS));
361# else
362 __asm
363 {
364 mov ax, cs
365 mov [SelCS], ax
366 }
367# endif
368 return SelCS;
369}
370#endif
371
372
373/**
374 * Get the DS register.
375 * @returns DS.
376 */
377#if RT_INLINE_ASM_EXTERNAL
378RT_ASM_DECL_PRAGMA_WATCOM(RTSEL) ASMGetDS(void);
379#else
380DECLINLINE(RTSEL) ASMGetDS(void)
381{
382 RTSEL SelDS;
383# if RT_INLINE_ASM_GNU_STYLE
384 __asm__ __volatile__("movw %%ds, %0\n\t" : "=r" (SelDS));
385# else
386 __asm
387 {
388 mov ax, ds
389 mov [SelDS], ax
390 }
391# endif
392 return SelDS;
393}
394#endif
395
396
397/**
398 * Get the ES register.
399 * @returns ES.
400 */
401#if RT_INLINE_ASM_EXTERNAL
402RT_ASM_DECL_PRAGMA_WATCOM(RTSEL) ASMGetES(void);
403#else
404DECLINLINE(RTSEL) ASMGetES(void)
405{
406 RTSEL SelES;
407# if RT_INLINE_ASM_GNU_STYLE
408 __asm__ __volatile__("movw %%es, %0\n\t" : "=r" (SelES));
409# else
410 __asm
411 {
412 mov ax, es
413 mov [SelES], ax
414 }
415# endif
416 return SelES;
417}
418#endif
419
420
421/**
422 * Get the FS register.
423 * @returns FS.
424 */
425#if RT_INLINE_ASM_EXTERNAL
426RT_ASM_DECL_PRAGMA_WATCOM(RTSEL) ASMGetFS(void);
427#else
428DECLINLINE(RTSEL) ASMGetFS(void)
429{
430 RTSEL SelFS;
431# if RT_INLINE_ASM_GNU_STYLE
432 __asm__ __volatile__("movw %%fs, %0\n\t" : "=r" (SelFS));
433# else
434 __asm
435 {
436 mov ax, fs
437 mov [SelFS], ax
438 }
439# endif
440 return SelFS;
441}
442# endif
443
444#ifdef RT_ARCH_AMD64
445
446/**
447 * Get the FS base register.
448 * @returns FS base address.
449 */
450#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2015 /*?*/
451DECLASM(uint64_t) ASMGetFSBase(void);
452#else
453DECLINLINE(uint64_t) ASMGetFSBase(void)
454{
455# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2015
456 return (uint64_t)_readfsbase_u64();
457# elif RT_INLINE_ASM_GNU_STYLE
458 uint64_t uFSBase;
459 __asm__ __volatile__("rdfsbase %0\n\t" : "=r" (uFSBase));
460 return uFSBase;
461# endif
462}
463# endif
464
465
466/**
467 * Set the FS base register.
468 * @param uNewBase The new base value.
469 */
470#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2015 /*?*/
471DECLASM(void) ASMSetFSBase(uint64_t uNewBase);
472#else
473DECLINLINE(void) ASMSetFSBase(uint64_t uNewBase)
474{
475# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2015
476 _writefsbase_u64(uNewBase);
477# elif RT_INLINE_ASM_GNU_STYLE
478 __asm__ __volatile__("wrfsbase %0\n\t" : : "r" (uNewBase));
479# endif
480}
481# endif
482
483#endif /* RT_ARCH_AMD64 */
484
485/**
486 * Get the GS register.
487 * @returns GS.
488 */
489#if RT_INLINE_ASM_EXTERNAL
490RT_ASM_DECL_PRAGMA_WATCOM(RTSEL) ASMGetGS(void);
491#else
492DECLINLINE(RTSEL) ASMGetGS(void)
493{
494 RTSEL SelGS;
495# if RT_INLINE_ASM_GNU_STYLE
496 __asm__ __volatile__("movw %%gs, %0\n\t" : "=r" (SelGS));
497# else
498 __asm
499 {
500 mov ax, gs
501 mov [SelGS], ax
502 }
503# endif
504 return SelGS;
505}
506#endif
507
508#ifdef RT_ARCH_AMD64
509
510/**
511 * Get the GS base register.
512 * @returns GS base address.
513 */
514#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2015 /*?*/
515DECLASM(uint64_t) ASMGetGSBase(void);
516#else
517DECLINLINE(uint64_t) ASMGetGSBase(void)
518{
519# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2015
520 return (uint64_t)_readgsbase_u64();
521# elif RT_INLINE_ASM_GNU_STYLE
522 uint64_t uGSBase;
523 __asm__ __volatile__("rdgsbase %0\n\t" : "=r" (uGSBase));
524 return uGSBase;
525# endif
526}
527# endif
528
529
530/**
531 * Set the GS base register.
532 * @param uNewBase The new base value.
533 */
534#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2015 /*?*/
535DECLASM(void) ASMSetGSBase(uint64_t uNewBase);
536#else
537DECLINLINE(void) ASMSetGSBase(uint64_t uNewBase)
538{
539# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2015
540 _writegsbase_u64(uNewBase);
541# elif RT_INLINE_ASM_GNU_STYLE
542 __asm__ __volatile__("wrgsbase %0\n\t" : : "r" (uNewBase));
543# endif
544}
545# endif
546
547#endif /* RT_ARCH_AMD64 */
548
549
550/**
551 * Get the SS register.
552 * @returns SS.
553 */
554#if RT_INLINE_ASM_EXTERNAL
555RT_ASM_DECL_PRAGMA_WATCOM(RTSEL) ASMGetSS(void);
556#else
557DECLINLINE(RTSEL) ASMGetSS(void)
558{
559 RTSEL SelSS;
560# if RT_INLINE_ASM_GNU_STYLE
561 __asm__ __volatile__("movw %%ss, %0\n\t" : "=r" (SelSS));
562# else
563 __asm
564 {
565 mov ax, ss
566 mov [SelSS], ax
567 }
568# endif
569 return SelSS;
570}
571#endif
572
573
574/**
575 * Get the TR register.
576 * @returns TR.
577 */
578#if RT_INLINE_ASM_EXTERNAL
579RT_ASM_DECL_PRAGMA_WATCOM(RTSEL) ASMGetTR(void);
580#else
581DECLINLINE(RTSEL) ASMGetTR(void)
582{
583 RTSEL SelTR;
584# if RT_INLINE_ASM_GNU_STYLE
585 __asm__ __volatile__("str %w0\n\t" : "=r" (SelTR));
586# else
587 __asm
588 {
589 str ax
590 mov [SelTR], ax
591 }
592# endif
593 return SelTR;
594}
595#endif
596
597
598/**
599 * Get the LDTR register.
600 * @returns LDTR.
601 */
602#if RT_INLINE_ASM_EXTERNAL
603RT_ASM_DECL_PRAGMA_WATCOM(RTSEL) ASMGetLDTR(void);
604#else
605DECLINLINE(RTSEL) ASMGetLDTR(void)
606{
607 RTSEL SelLDTR;
608# if RT_INLINE_ASM_GNU_STYLE
609 __asm__ __volatile__("sldt %w0\n\t" : "=r" (SelLDTR));
610# else
611 __asm
612 {
613 sldt ax
614 mov [SelLDTR], ax
615 }
616# endif
617 return SelLDTR;
618}
619#endif
620
621
622/**
623 * Get the access rights for the segment selector.
624 *
625 * @returns The access rights on success or UINT32_MAX on failure.
626 * @param uSel The selector value.
627 *
628 * @remarks Using UINT32_MAX for failure is chosen because valid access rights
629 * always have bits 0:7 as 0 (on both Intel & AMD).
630 */
631#if RT_INLINE_ASM_EXTERNAL
632RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMGetSegAttr(uint32_t uSel);
633#else
634DECLINLINE(uint32_t) ASMGetSegAttr(uint32_t uSel)
635{
636 uint32_t uAttr;
637 /* LAR only accesses 16-bit of the source operand, but eax for the
638 destination operand is required for getting the full 32-bit access rights. */
639# if RT_INLINE_ASM_GNU_STYLE
640 __asm__ __volatile__("lar %1, %%eax\n\t"
641 "jz done%=\n\t"
642 "movl $0xffffffff, %%eax\n\t"
643 "done%=:\n\t"
644 "movl %%eax, %0\n\t"
645 : "=r" (uAttr)
646 : "r" (uSel)
647 : "cc", "%eax");
648# else
649 __asm
650 {
651 lar eax, [uSel]
652 jz done
653 mov eax, 0ffffffffh
654 done:
655 mov [uAttr], eax
656 }
657# endif
658 return uAttr;
659}
660#endif
661
662
663/**
664 * Get the [RE]FLAGS register.
665 * @returns [RE]FLAGS.
666 */
667#if RT_INLINE_ASM_EXTERNAL /*&& RT_INLINE_ASM_USES_INTRIN < 15 - buggy intrinsics in VC++ 2010, reordering/optimizers issues. */
668RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTREG) ASMGetFlags(void);
669#else
670DECLINLINE(RTCCUINTREG) ASMGetFlags(void)
671{
672 RTCCUINTREG uFlags;
673# if RT_INLINE_ASM_GNU_STYLE
674# ifdef RT_ARCH_AMD64
675 __asm__ __volatile__("pushfq\n\t"
676 "popq %0\n\t"
677 : "=r" (uFlags));
678# else
679 __asm__ __volatile__("pushfl\n\t"
680 "popl %0\n\t"
681 : "=r" (uFlags));
682# endif
683# elif RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2008
684 uFlags = __readeflags();
685# else
686 __asm
687 {
688# ifdef RT_ARCH_AMD64
689 pushfq
690 pop [uFlags]
691# else
692 pushfd
693 pop [uFlags]
694# endif
695 }
696# endif
697 return uFlags;
698}
699#endif
700
701
702/**
703 * Set the [RE]FLAGS register.
704 * @param uFlags The new [RE]FLAGS value.
705 */
706#if RT_INLINE_ASM_EXTERNAL /*&& RT_INLINE_ASM_USES_INTRIN < 15 - see __readeflags() above. */
707RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetFlags(RTCCUINTREG uFlags);
708#else
709DECLINLINE(void) ASMSetFlags(RTCCUINTREG uFlags)
710{
711# if RT_INLINE_ASM_GNU_STYLE
712# ifdef RT_ARCH_AMD64
713 __asm__ __volatile__("pushq %0\n\t"
714 "popfq\n\t"
715 : : "g" (uFlags));
716# else
717 __asm__ __volatile__("pushl %0\n\t"
718 "popfl\n\t"
719 : : "g" (uFlags));
720# endif
721# elif RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2008
722 __writeeflags(uFlags);
723# else
724 __asm
725 {
726# ifdef RT_ARCH_AMD64
727 push [uFlags]
728 popfq
729# else
730 push [uFlags]
731 popfd
732# endif
733 }
734# endif
735}
736#endif
737
738
739/**
740 * Modifies the [RE]FLAGS register.
741 * @returns Original value.
742 * @param fAndEfl Flags to keep (applied first).
743 * @param fOrEfl Flags to be set.
744 */
745#if RT_INLINE_ASM_EXTERNAL /*&& RT_INLINE_ASM_USES_INTRIN < 15 - buggy intrinsics in VC++ 2010, reordering/optimizers issues. */
746RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTREG) ASMChangeFlags(RTCCUINTREG fAndEfl, RTCCUINTREG fOrEfl);
747#else
748DECLINLINE(RTCCUINTREG) ASMChangeFlags(RTCCUINTREG fAndEfl, RTCCUINTREG fOrEfl)
749{
750 RTCCUINTREG fOldEfl;
751# if RT_INLINE_ASM_GNU_STYLE
752# ifdef RT_ARCH_AMD64
753 __asm__ __volatile__("pushfq\n\t"
754 "movq (%%rsp), %0\n\t"
755 "andq %0, %1\n\t"
756 "orq %3, %1\n\t"
757 "mov %1, (%%rsp)\n\t"
758 "popfq\n\t"
759 : "=&r" (fOldEfl),
760 "=r" (fAndEfl)
761 : "1" (fAndEfl),
762 "rn" (fOrEfl) );
763# else
764 __asm__ __volatile__("pushfl\n\t"
765 "movl (%%esp), %0\n\t"
766 "andl %1, (%%esp)\n\t"
767 "orl %2, (%%esp)\n\t"
768 "popfl\n\t"
769 : "=&r" (fOldEfl)
770 : "rn" (fAndEfl),
771 "rn" (fOrEfl) );
772# endif
773# elif RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2008
774 fOldEfl = __readeflags();
775 __writeeflags((fOldEfl & fAndEfl) | fOrEfl);
776# else
777 __asm
778 {
779# ifdef RT_ARCH_AMD64
780 mov rdx, [fAndEfl]
781 mov rcx, [fOrEfl]
782 pushfq
783 mov rax, [rsp]
784 and rdx, rax
785 or rdx, rcx
786 mov [rsp], rdx
787 popfq
788 mov [fOldEfl], rax
789# else
790 mov edx, [fAndEfl]
791 mov ecx, [fOrEfl]
792 pushfd
793 mov eax, [esp]
794 and edx, eax
795 or edx, ecx
796 mov [esp], edx
797 popfd
798 mov [fOldEfl], eax
799# endif
800 }
801# endif
802 return fOldEfl;
803}
804#endif
805
806
807/**
808 * Modifies the [RE]FLAGS register by ORing in one or more flags.
809 * @returns Original value.
810 * @param fOrEfl The flags to be set (ORed in).
811 */
812#if RT_INLINE_ASM_EXTERNAL /*&& RT_INLINE_ASM_USES_INTRIN < 15 - buggy intrinsics in VC++ 2010, reordering/optimizers issues. */
813RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTREG) ASMAddFlags(RTCCUINTREG fOrEfl);
814#else
815DECLINLINE(RTCCUINTREG) ASMAddFlags(RTCCUINTREG fOrEfl)
816{
817 RTCCUINTREG fOldEfl;
818# if RT_INLINE_ASM_GNU_STYLE
819# ifdef RT_ARCH_AMD64
820 __asm__ __volatile__("pushfq\n\t"
821 "movq (%%rsp), %0\n\t"
822 "orq %1, (%%rsp)\n\t"
823 "popfq\n\t"
824 : "=&r" (fOldEfl)
825 : "rn" (fOrEfl) );
826# else
827 __asm__ __volatile__("pushfl\n\t"
828 "movl (%%esp), %0\n\t"
829 "orl %1, (%%esp)\n\t"
830 "popfl\n\t"
831 : "=&r" (fOldEfl)
832 : "rn" (fOrEfl) );
833# endif
834# elif RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2008
835 fOldEfl = __readeflags();
836 __writeeflags(fOldEfl | fOrEfl);
837# else
838 __asm
839 {
840# ifdef RT_ARCH_AMD64
841 mov rcx, [fOrEfl]
842 pushfq
843 mov rdx, [rsp]
844 or [rsp], rcx
845 popfq
846 mov [fOldEfl], rax
847# else
848 mov ecx, [fOrEfl]
849 pushfd
850 mov edx, [esp]
851 or [esp], ecx
852 popfd
853 mov [fOldEfl], eax
854# endif
855 }
856# endif
857 return fOldEfl;
858}
859#endif
860
861
862/**
863 * Modifies the [RE]FLAGS register by AND'ing out one or more flags.
864 * @returns Original value.
865 * @param fAndEfl The flags to keep.
866 */
867#if RT_INLINE_ASM_EXTERNAL /*&& RT_INLINE_ASM_USES_INTRIN < 15 - buggy intrinsics in VC++ 2010, reordering/optimizers issues. */
868RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTREG) ASMClearFlags(RTCCUINTREG fAndEfl);
869#else
870DECLINLINE(RTCCUINTREG) ASMClearFlags(RTCCUINTREG fAndEfl)
871{
872 RTCCUINTREG fOldEfl;
873# if RT_INLINE_ASM_GNU_STYLE
874# ifdef RT_ARCH_AMD64
875 __asm__ __volatile__("pushfq\n\t"
876 "movq (%%rsp), %0\n\t"
877 "andq %1, (%%rsp)\n\t"
878 "popfq\n\t"
879 : "=&r" (fOldEfl)
880 : "rn" (fAndEfl) );
881# else
882 __asm__ __volatile__("pushfl\n\t"
883 "movl (%%esp), %0\n\t"
884 "andl %1, (%%esp)\n\t"
885 "popfl\n\t"
886 : "=&r" (fOldEfl)
887 : "rn" (fAndEfl) );
888# endif
889# elif RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2008
890 fOldEfl = __readeflags();
891 __writeeflags(fOldEfl & fAndEfl);
892# else
893 __asm
894 {
895# ifdef RT_ARCH_AMD64
896 mov rdx, [fAndEfl]
897 pushfq
898 mov rdx, [rsp]
899 and [rsp], rdx
900 popfq
901 mov [fOldEfl], rax
902# else
903 mov edx, [fAndEfl]
904 pushfd
905 mov edx, [esp]
906 and [esp], edx
907 popfd
908 mov [fOldEfl], eax
909# endif
910 }
911# endif
912 return fOldEfl;
913}
914#endif
915
916
917/**
918 * Gets the content of the CPU timestamp counter register.
919 *
920 * @returns TSC.
921 */
922#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
923RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMReadTSC(void);
924#else
925DECLINLINE(uint64_t) ASMReadTSC(void)
926{
927 RTUINT64U u;
928# if RT_INLINE_ASM_GNU_STYLE
929 __asm__ __volatile__("rdtsc\n\t" : "=a" (u.s.Lo), "=d" (u.s.Hi));
930# else
931# if RT_INLINE_ASM_USES_INTRIN
932 u.u = __rdtsc();
933# else
934 __asm
935 {
936 rdtsc
937 mov [u.s.Lo], eax
938 mov [u.s.Hi], edx
939 }
940# endif
941# endif
942 return u.u;
943}
944#endif
945
946
947/**
948 * Gets the content of the CPU timestamp counter register and the
949 * assoicated AUX value.
950 *
951 * @returns TSC.
952 * @param puAux Where to store the AUX value.
953 */
954#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2008
955RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMReadTscWithAux(uint32_t RT_FAR *puAux);
956#else
957DECLINLINE(uint64_t) ASMReadTscWithAux(uint32_t RT_FAR *puAux)
958{
959 RTUINT64U u;
960# if RT_INLINE_ASM_GNU_STYLE
961 /* rdtscp is not supported by ancient linux build VM of course :-( */
962 /*__asm__ __volatile__("rdtscp\n\t" : "=a" (u.s.Lo), "=d" (u.s.Hi), "=c" (*puAux)); */
963 __asm__ __volatile__(".byte 0x0f,0x01,0xf9\n\t" : "=a" (u.s.Lo), "=d" (u.s.Hi), "=c" (*puAux));
964# else
965# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2008
966 u.u = __rdtscp(puAux);
967# else
968 __asm
969 {
970 rdtscp
971 mov [u.s.Lo], eax
972 mov [u.s.Hi], edx
973 mov eax, [puAux]
974 mov [eax], ecx
975 }
976# endif
977# endif
978 return u.u;
979}
980#endif
981
982
983/**
984 * Performs the cpuid instruction returning all registers.
985 *
986 * @param uOperator CPUID operation (eax).
987 * @param pvEAX Where to store eax.
988 * @param pvEBX Where to store ebx.
989 * @param pvECX Where to store ecx.
990 * @param pvEDX Where to store edx.
991 * @remark We're using void pointers to ease the use of special bitfield structures and such.
992 */
993#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
994DECLASM(void) ASMCpuId(uint32_t uOperator, void RT_FAR *pvEAX, void RT_FAR *pvEBX, void RT_FAR *pvECX, void RT_FAR *pvEDX);
995#else
996DECLINLINE(void) ASMCpuId(uint32_t uOperator, void RT_FAR *pvEAX, void RT_FAR *pvEBX, void RT_FAR *pvECX, void RT_FAR *pvEDX)
997{
998# if RT_INLINE_ASM_GNU_STYLE
999# ifdef RT_ARCH_AMD64
1000 RTCCUINTREG uRAX, uRBX, uRCX, uRDX;
1001 __asm__ __volatile__ ("cpuid\n\t"
1002 : "=a" (uRAX),
1003 "=b" (uRBX),
1004 "=c" (uRCX),
1005 "=d" (uRDX)
1006 : "0" (uOperator), "2" (0));
1007 *(uint32_t RT_FAR *)pvEAX = (uint32_t)uRAX;
1008 *(uint32_t RT_FAR *)pvEBX = (uint32_t)uRBX;
1009 *(uint32_t RT_FAR *)pvECX = (uint32_t)uRCX;
1010 *(uint32_t RT_FAR *)pvEDX = (uint32_t)uRDX;
1011# else
1012 __asm__ __volatile__ ("xchgl %%ebx, %1\n\t"
1013 "cpuid\n\t"
1014 "xchgl %%ebx, %1\n\t"
1015 : "=a" (*(uint32_t *)pvEAX),
1016 "=r" (*(uint32_t *)pvEBX),
1017 "=c" (*(uint32_t *)pvECX),
1018 "=d" (*(uint32_t *)pvEDX)
1019 : "0" (uOperator), "2" (0));
1020# endif
1021
1022# elif RT_INLINE_ASM_USES_INTRIN
1023 int aInfo[4];
1024 __cpuid(aInfo, uOperator);
1025 *(uint32_t RT_FAR *)pvEAX = aInfo[0];
1026 *(uint32_t RT_FAR *)pvEBX = aInfo[1];
1027 *(uint32_t RT_FAR *)pvECX = aInfo[2];
1028 *(uint32_t RT_FAR *)pvEDX = aInfo[3];
1029
1030# else
1031 uint32_t uEAX;
1032 uint32_t uEBX;
1033 uint32_t uECX;
1034 uint32_t uEDX;
1035 __asm
1036 {
1037 push ebx
1038 mov eax, [uOperator]
1039 cpuid
1040 mov [uEAX], eax
1041 mov [uEBX], ebx
1042 mov [uECX], ecx
1043 mov [uEDX], edx
1044 pop ebx
1045 }
1046 *(uint32_t RT_FAR *)pvEAX = uEAX;
1047 *(uint32_t RT_FAR *)pvEBX = uEBX;
1048 *(uint32_t RT_FAR *)pvECX = uECX;
1049 *(uint32_t RT_FAR *)pvEDX = uEDX;
1050# endif
1051}
1052#endif
1053
1054
1055/**
1056 * Performs the CPUID instruction with EAX and ECX input returning ALL output
1057 * registers.
1058 *
1059 * @param uOperator CPUID operation (eax).
1060 * @param uIdxECX ecx index
1061 * @param pvEAX Where to store eax.
1062 * @param pvEBX Where to store ebx.
1063 * @param pvECX Where to store ecx.
1064 * @param pvEDX Where to store edx.
1065 * @remark We're using void pointers to ease the use of special bitfield structures and such.
1066 */
1067#if RT_INLINE_ASM_EXTERNAL || RT_INLINE_ASM_USES_INTRIN
1068DECLASM(void) ASMCpuId_Idx_ECX(uint32_t uOperator, uint32_t uIdxECX, void RT_FAR *pvEAX, void RT_FAR *pvEBX, void RT_FAR *pvECX, void RT_FAR *pvEDX);
1069#else
1070DECLINLINE(void) ASMCpuId_Idx_ECX(uint32_t uOperator, uint32_t uIdxECX, void RT_FAR *pvEAX, void RT_FAR *pvEBX, void RT_FAR *pvECX, void RT_FAR *pvEDX)
1071{
1072# if RT_INLINE_ASM_GNU_STYLE
1073# ifdef RT_ARCH_AMD64
1074 RTCCUINTREG uRAX, uRBX, uRCX, uRDX;
1075 __asm__ ("cpuid\n\t"
1076 : "=a" (uRAX),
1077 "=b" (uRBX),
1078 "=c" (uRCX),
1079 "=d" (uRDX)
1080 : "0" (uOperator),
1081 "2" (uIdxECX));
1082 *(uint32_t RT_FAR *)pvEAX = (uint32_t)uRAX;
1083 *(uint32_t RT_FAR *)pvEBX = (uint32_t)uRBX;
1084 *(uint32_t RT_FAR *)pvECX = (uint32_t)uRCX;
1085 *(uint32_t RT_FAR *)pvEDX = (uint32_t)uRDX;
1086# else
1087 __asm__ ("xchgl %%ebx, %1\n\t"
1088 "cpuid\n\t"
1089 "xchgl %%ebx, %1\n\t"
1090 : "=a" (*(uint32_t *)pvEAX),
1091 "=r" (*(uint32_t *)pvEBX),
1092 "=c" (*(uint32_t *)pvECX),
1093 "=d" (*(uint32_t *)pvEDX)
1094 : "0" (uOperator),
1095 "2" (uIdxECX));
1096# endif
1097
1098# elif RT_INLINE_ASM_USES_INTRIN
1099 int aInfo[4];
1100 __cpuidex(aInfo, uOperator, uIdxECX);
1101 *(uint32_t RT_FAR *)pvEAX = aInfo[0];
1102 *(uint32_t RT_FAR *)pvEBX = aInfo[1];
1103 *(uint32_t RT_FAR *)pvECX = aInfo[2];
1104 *(uint32_t RT_FAR *)pvEDX = aInfo[3];
1105
1106# else
1107 uint32_t uEAX;
1108 uint32_t uEBX;
1109 uint32_t uECX;
1110 uint32_t uEDX;
1111 __asm
1112 {
1113 push ebx
1114 mov eax, [uOperator]
1115 mov ecx, [uIdxECX]
1116 cpuid
1117 mov [uEAX], eax
1118 mov [uEBX], ebx
1119 mov [uECX], ecx
1120 mov [uEDX], edx
1121 pop ebx
1122 }
1123 *(uint32_t RT_FAR *)pvEAX = uEAX;
1124 *(uint32_t RT_FAR *)pvEBX = uEBX;
1125 *(uint32_t RT_FAR *)pvECX = uECX;
1126 *(uint32_t RT_FAR *)pvEDX = uEDX;
1127# endif
1128}
1129#endif
1130
1131
1132/**
1133 * CPUID variant that initializes all 4 registers before the CPUID instruction.
1134 *
1135 * @returns The EAX result value.
1136 * @param uOperator CPUID operation (eax).
1137 * @param uInitEBX The value to assign EBX prior to the CPUID instruction.
1138 * @param uInitECX The value to assign ECX prior to the CPUID instruction.
1139 * @param uInitEDX The value to assign EDX prior to the CPUID instruction.
1140 * @param pvEAX Where to store eax. Optional.
1141 * @param pvEBX Where to store ebx. Optional.
1142 * @param pvECX Where to store ecx. Optional.
1143 * @param pvEDX Where to store edx. Optional.
1144 */
1145DECLASM(uint32_t) ASMCpuIdExSlow(uint32_t uOperator, uint32_t uInitEBX, uint32_t uInitECX, uint32_t uInitEDX,
1146 void RT_FAR *pvEAX, void RT_FAR *pvEBX, void RT_FAR *pvECX, void RT_FAR *pvEDX);
1147
1148
1149/**
1150 * Performs the cpuid instruction returning ecx and edx.
1151 *
1152 * @param uOperator CPUID operation (eax).
1153 * @param pvECX Where to store ecx.
1154 * @param pvEDX Where to store edx.
1155 * @remark We're using void pointers to ease the use of special bitfield structures and such.
1156 */
1157#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1158RT_ASM_DECL_PRAGMA_WATCOM(void) ASMCpuId_ECX_EDX(uint32_t uOperator, void RT_FAR *pvECX, void RT_FAR *pvEDX);
1159#else
1160DECLINLINE(void) ASMCpuId_ECX_EDX(uint32_t uOperator, void RT_FAR *pvECX, void RT_FAR *pvEDX)
1161{
1162 uint32_t uEBX;
1163 ASMCpuId(uOperator, &uOperator, &uEBX, pvECX, pvEDX);
1164}
1165#endif
1166
1167
1168/**
1169 * Performs the cpuid instruction returning eax.
1170 *
1171 * @param uOperator CPUID operation (eax).
1172 * @returns EAX after cpuid operation.
1173 */
1174#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1175RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMCpuId_EAX(uint32_t uOperator);
1176#else
1177DECLINLINE(uint32_t) ASMCpuId_EAX(uint32_t uOperator)
1178{
1179 RTCCUINTREG xAX;
1180# if RT_INLINE_ASM_GNU_STYLE
1181# ifdef RT_ARCH_AMD64
1182 __asm__ ("cpuid"
1183 : "=a" (xAX)
1184 : "0" (uOperator)
1185 : "rbx", "rcx", "rdx");
1186# elif (defined(PIC) || defined(__PIC__)) && defined(__i386__)
1187 __asm__ ("push %%ebx\n\t"
1188 "cpuid\n\t"
1189 "pop %%ebx\n\t"
1190 : "=a" (xAX)
1191 : "0" (uOperator)
1192 : "ecx", "edx");
1193# else
1194 __asm__ ("cpuid"
1195 : "=a" (xAX)
1196 : "0" (uOperator)
1197 : "edx", "ecx", "ebx");
1198# endif
1199
1200# elif RT_INLINE_ASM_USES_INTRIN
1201 int aInfo[4];
1202 __cpuid(aInfo, uOperator);
1203 xAX = aInfo[0];
1204
1205# else
1206 __asm
1207 {
1208 push ebx
1209 mov eax, [uOperator]
1210 cpuid
1211 mov [xAX], eax
1212 pop ebx
1213 }
1214# endif
1215 return (uint32_t)xAX;
1216}
1217#endif
1218
1219
1220/**
1221 * Performs the cpuid instruction returning ebx.
1222 *
1223 * @param uOperator CPUID operation (eax).
1224 * @returns EBX after cpuid operation.
1225 */
1226#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1227RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMCpuId_EBX(uint32_t uOperator);
1228#else
1229DECLINLINE(uint32_t) ASMCpuId_EBX(uint32_t uOperator)
1230{
1231 RTCCUINTREG xBX;
1232# if RT_INLINE_ASM_GNU_STYLE
1233# ifdef RT_ARCH_AMD64
1234 RTCCUINTREG uSpill;
1235 __asm__ ("cpuid"
1236 : "=a" (uSpill),
1237 "=b" (xBX)
1238 : "0" (uOperator)
1239 : "rdx", "rcx");
1240# elif (defined(PIC) || defined(__PIC__)) && defined(__i386__)
1241 __asm__ ("push %%ebx\n\t"
1242 "cpuid\n\t"
1243 "mov %%ebx, %%edx\n\t"
1244 "pop %%ebx\n\t"
1245 : "=a" (uOperator),
1246 "=d" (xBX)
1247 : "0" (uOperator)
1248 : "ecx");
1249# else
1250 __asm__ ("cpuid"
1251 : "=a" (uOperator),
1252 "=b" (xBX)
1253 : "0" (uOperator)
1254 : "edx", "ecx");
1255# endif
1256
1257# elif RT_INLINE_ASM_USES_INTRIN
1258 int aInfo[4];
1259 __cpuid(aInfo, uOperator);
1260 xBX = aInfo[1];
1261
1262# else
1263 __asm
1264 {
1265 push ebx
1266 mov eax, [uOperator]
1267 cpuid
1268 mov [xBX], ebx
1269 pop ebx
1270 }
1271# endif
1272 return (uint32_t)xBX;
1273}
1274#endif
1275
1276
1277/**
1278 * Performs the cpuid instruction returning ecx.
1279 *
1280 * @param uOperator CPUID operation (eax).
1281 * @returns ECX after cpuid operation.
1282 */
1283#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1284RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMCpuId_ECX(uint32_t uOperator);
1285#else
1286DECLINLINE(uint32_t) ASMCpuId_ECX(uint32_t uOperator)
1287{
1288 RTCCUINTREG xCX;
1289# if RT_INLINE_ASM_GNU_STYLE
1290# ifdef RT_ARCH_AMD64
1291 RTCCUINTREG uSpill;
1292 __asm__ ("cpuid"
1293 : "=a" (uSpill),
1294 "=c" (xCX)
1295 : "0" (uOperator)
1296 : "rbx", "rdx");
1297# elif (defined(PIC) || defined(__PIC__)) && defined(__i386__)
1298 __asm__ ("push %%ebx\n\t"
1299 "cpuid\n\t"
1300 "pop %%ebx\n\t"
1301 : "=a" (uOperator),
1302 "=c" (xCX)
1303 : "0" (uOperator)
1304 : "edx");
1305# else
1306 __asm__ ("cpuid"
1307 : "=a" (uOperator),
1308 "=c" (xCX)
1309 : "0" (uOperator)
1310 : "ebx", "edx");
1311
1312# endif
1313
1314# elif RT_INLINE_ASM_USES_INTRIN
1315 int aInfo[4];
1316 __cpuid(aInfo, uOperator);
1317 xCX = aInfo[2];
1318
1319# else
1320 __asm
1321 {
1322 push ebx
1323 mov eax, [uOperator]
1324 cpuid
1325 mov [xCX], ecx
1326 pop ebx
1327 }
1328# endif
1329 return (uint32_t)xCX;
1330}
1331#endif
1332
1333
1334/**
1335 * Performs the cpuid instruction returning edx.
1336 *
1337 * @param uOperator CPUID operation (eax).
1338 * @returns EDX after cpuid operation.
1339 */
1340#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1341RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMCpuId_EDX(uint32_t uOperator);
1342#else
1343DECLINLINE(uint32_t) ASMCpuId_EDX(uint32_t uOperator)
1344{
1345 RTCCUINTREG xDX;
1346# if RT_INLINE_ASM_GNU_STYLE
1347# ifdef RT_ARCH_AMD64
1348 RTCCUINTREG uSpill;
1349 __asm__ ("cpuid"
1350 : "=a" (uSpill),
1351 "=d" (xDX)
1352 : "0" (uOperator)
1353 : "rbx", "rcx");
1354# elif (defined(PIC) || defined(__PIC__)) && defined(__i386__)
1355 __asm__ ("push %%ebx\n\t"
1356 "cpuid\n\t"
1357 "pop %%ebx\n\t"
1358 : "=a" (uOperator),
1359 "=d" (xDX)
1360 : "0" (uOperator)
1361 : "ecx");
1362# else
1363 __asm__ ("cpuid"
1364 : "=a" (uOperator),
1365 "=d" (xDX)
1366 : "0" (uOperator)
1367 : "ebx", "ecx");
1368# endif
1369
1370# elif RT_INLINE_ASM_USES_INTRIN
1371 int aInfo[4];
1372 __cpuid(aInfo, uOperator);
1373 xDX = aInfo[3];
1374
1375# else
1376 __asm
1377 {
1378 push ebx
1379 mov eax, [uOperator]
1380 cpuid
1381 mov [xDX], edx
1382 pop ebx
1383 }
1384# endif
1385 return (uint32_t)xDX;
1386}
1387#endif
1388
1389
1390/**
1391 * Checks if the current CPU supports CPUID.
1392 *
1393 * @returns true if CPUID is supported.
1394 */
1395#ifdef __WATCOMC__
1396DECLASM(bool) ASMHasCpuId(void);
1397#else
1398DECLINLINE(bool) ASMHasCpuId(void)
1399{
1400# ifdef RT_ARCH_AMD64
1401 return true; /* ASSUME that all amd64 compatible CPUs have cpuid. */
1402# else /* !RT_ARCH_AMD64 */
1403 bool fRet = false;
1404# if RT_INLINE_ASM_GNU_STYLE
1405 uint32_t u1;
1406 uint32_t u2;
1407 __asm__ ("pushf\n\t"
1408 "pop %1\n\t"
1409 "mov %1, %2\n\t"
1410 "xorl $0x200000, %1\n\t"
1411 "push %1\n\t"
1412 "popf\n\t"
1413 "pushf\n\t"
1414 "pop %1\n\t"
1415 "cmpl %1, %2\n\t"
1416 "setne %0\n\t"
1417 "push %2\n\t"
1418 "popf\n\t"
1419 : "=m" (fRet), "=r" (u1), "=r" (u2));
1420# else
1421 __asm
1422 {
1423 pushfd
1424 pop eax
1425 mov ebx, eax
1426 xor eax, 0200000h
1427 push eax
1428 popfd
1429 pushfd
1430 pop eax
1431 cmp eax, ebx
1432 setne fRet
1433 push ebx
1434 popfd
1435 }
1436# endif
1437 return fRet;
1438# endif /* !RT_ARCH_AMD64 */
1439}
1440#endif
1441
1442
1443/**
1444 * Gets the APIC ID of the current CPU.
1445 *
1446 * @returns the APIC ID.
1447 */
1448#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1449RT_ASM_DECL_PRAGMA_WATCOM(uint8_t) ASMGetApicId(void);
1450#else
1451DECLINLINE(uint8_t) ASMGetApicId(void)
1452{
1453 RTCCUINTREG xBX;
1454# if RT_INLINE_ASM_GNU_STYLE
1455# ifdef RT_ARCH_AMD64
1456 RTCCUINTREG uSpill;
1457 __asm__ __volatile__ ("cpuid"
1458 : "=a" (uSpill),
1459 "=b" (xBX)
1460 : "0" (1)
1461 : "rcx", "rdx");
1462# elif (defined(PIC) || defined(__PIC__)) && defined(__i386__)
1463 RTCCUINTREG uSpill;
1464 __asm__ __volatile__ ("mov %%ebx,%1\n\t"
1465 "cpuid\n\t"
1466 "xchgl %%ebx,%1\n\t"
1467 : "=a" (uSpill),
1468 "=rm" (xBX)
1469 : "0" (1)
1470 : "ecx", "edx");
1471# else
1472 RTCCUINTREG uSpill;
1473 __asm__ __volatile__ ("cpuid"
1474 : "=a" (uSpill),
1475 "=b" (xBX)
1476 : "0" (1)
1477 : "ecx", "edx");
1478# endif
1479
1480# elif RT_INLINE_ASM_USES_INTRIN
1481 int aInfo[4];
1482 __cpuid(aInfo, 1);
1483 xBX = aInfo[1];
1484
1485# else
1486 __asm
1487 {
1488 push ebx
1489 mov eax, 1
1490 cpuid
1491 mov [xBX], ebx
1492 pop ebx
1493 }
1494# endif
1495 return (uint8_t)(xBX >> 24);
1496}
1497#endif
1498
1499
1500/**
1501 * Gets the APIC ID of the current CPU using leaf 0xb.
1502 *
1503 * @returns the APIC ID.
1504 */
1505#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2010 /*?*/
1506RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMGetApicIdExt0B(void);
1507#else
1508DECLINLINE(uint32_t) ASMGetApicIdExt0B(void)
1509{
1510# if RT_INLINE_ASM_GNU_STYLE
1511 RTCCUINTREG xDX;
1512# ifdef RT_ARCH_AMD64
1513 RTCCUINTREG uSpillEax, uSpillEcx;
1514 __asm__ __volatile__ ("cpuid"
1515 : "=a" (uSpillEax),
1516 "=c" (uSpillEcx),
1517 "=d" (xDX)
1518 : "0" (0xb),
1519 "1" (0)
1520 : "rbx");
1521# elif (defined(PIC) || defined(__PIC__)) && defined(__i386__)
1522 RTCCUINTREG uSpillEax, uSpillEcx, uSpillEbx;
1523 __asm__ __volatile__ ("mov %%ebx,%2\n\t"
1524 "cpuid\n\t"
1525 "xchgl %%ebx,%2\n\t"
1526 : "=a" (uSpillEax),
1527 "=c" (uSpillEcx),
1528 "=rm" (uSpillEbx),
1529 "=d" (xDX)
1530 : "0" (0xb),
1531 "1" (0));
1532# else
1533 RTCCUINTREG uSpillEax, uSpillEcx;
1534 __asm__ __volatile__ ("cpuid"
1535 : "=a" (uSpillEax),
1536 "=c" (uSpillEcx),
1537 "=d" (xDX)
1538 : "0" (0xb),
1539 "1" (0)
1540 : "ebx");
1541# endif
1542 return (uint32_t)xDX;
1543
1544# elif RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2010 /*?*/
1545
1546 int aInfo[4];
1547 __cpuidex(aInfo, 0xb, 0);
1548 return aInfo[3];
1549
1550# else
1551 RTCCUINTREG xDX;
1552 __asm
1553 {
1554 push ebx
1555 mov eax, 0xb
1556 xor ecx, ecx
1557 cpuid
1558 mov [xDX], edx
1559 pop ebx
1560 }
1561 return (uint32_t)xDX;
1562# endif
1563}
1564#endif
1565
1566
1567/**
1568 * Gets the APIC ID of the current CPU using leaf 8000001E.
1569 *
1570 * @returns the APIC ID.
1571 */
1572DECLINLINE(uint32_t) ASMGetApicIdExt8000001E(void)
1573{
1574 return ASMCpuId_EAX(0x8000001e);
1575}
1576
1577
1578/**
1579 * Tests if this is a genuine Intel CPU.
1580 *
1581 * @returns true/false.
1582 * @remarks ASSUMES that cpuid is supported by the CPU.
1583 */
1584DECLINLINE(bool) ASMIsIntelCpu(void)
1585{
1586 uint32_t uEAX, uEBX, uECX, uEDX;
1587 ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
1588 return RTX86IsIntelCpu(uEBX, uECX, uEDX);
1589}
1590
1591
1592/**
1593 * Tests if this is an authentic AMD CPU.
1594 *
1595 * @returns true/false.
1596 * @remarks ASSUMES that cpuid is supported by the CPU.
1597 */
1598DECLINLINE(bool) ASMIsAmdCpu(void)
1599{
1600 uint32_t uEAX, uEBX, uECX, uEDX;
1601 ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
1602 return RTX86IsAmdCpu(uEBX, uECX, uEDX);
1603}
1604
1605
1606/**
1607 * Tests if this is a centaur hauling VIA CPU.
1608 *
1609 * @returns true/false.
1610 * @remarks ASSUMES that cpuid is supported by the CPU.
1611 */
1612DECLINLINE(bool) ASMIsViaCentaurCpu(void)
1613{
1614 uint32_t uEAX, uEBX, uECX, uEDX;
1615 ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
1616 return RTX86IsViaCentaurCpu(uEBX, uECX, uEDX);
1617}
1618
1619
1620/**
1621 * Tests if this is a Shanghai CPU.
1622 *
1623 * @returns true/false.
1624 * @remarks ASSUMES that cpuid is supported by the CPU.
1625 */
1626DECLINLINE(bool) ASMIsShanghaiCpu(void)
1627{
1628 uint32_t uEAX, uEBX, uECX, uEDX;
1629 ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
1630 return RTX86IsShanghaiCpu(uEBX, uECX, uEDX);
1631}
1632
1633
1634/**
1635 * Tests if this is a genuine Hygon CPU.
1636 *
1637 * @returns true/false.
1638 * @remarks ASSUMES that cpuid is supported by the CPU.
1639 */
1640DECLINLINE(bool) ASMIsHygonCpu(void)
1641{
1642 uint32_t uEAX, uEBX, uECX, uEDX;
1643 ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
1644 return RTX86IsHygonCpu(uEBX, uECX, uEDX);
1645}
1646
1647
1648/**
1649 * Get cr0.
1650 * @returns cr0.
1651 */
1652#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1653RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTXREG) ASMGetCR0(void);
1654#else
1655DECLINLINE(RTCCUINTXREG) ASMGetCR0(void)
1656{
1657 RTCCUINTXREG uCR0;
1658# if RT_INLINE_ASM_USES_INTRIN
1659 uCR0 = __readcr0();
1660
1661# elif RT_INLINE_ASM_GNU_STYLE
1662# ifdef RT_ARCH_AMD64
1663 __asm__ __volatile__("movq %%cr0, %0\t\n" : "=r" (uCR0));
1664# else
1665 __asm__ __volatile__("movl %%cr0, %0\t\n" : "=r" (uCR0));
1666# endif
1667# else
1668 __asm
1669 {
1670# ifdef RT_ARCH_AMD64
1671 mov rax, cr0
1672 mov [uCR0], rax
1673# else
1674 mov eax, cr0
1675 mov [uCR0], eax
1676# endif
1677 }
1678# endif
1679 return uCR0;
1680}
1681#endif
1682
1683
1684/**
1685 * Sets the CR0 register.
1686 * @param uCR0 The new CR0 value.
1687 */
1688#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1689RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetCR0(RTCCUINTXREG uCR0);
1690#else
1691DECLINLINE(void) ASMSetCR0(RTCCUINTXREG uCR0)
1692{
1693# if RT_INLINE_ASM_USES_INTRIN
1694 __writecr0(uCR0);
1695
1696# elif RT_INLINE_ASM_GNU_STYLE
1697# ifdef RT_ARCH_AMD64
1698 __asm__ __volatile__("movq %0, %%cr0\n\t" :: "r" (uCR0));
1699# else
1700 __asm__ __volatile__("movl %0, %%cr0\n\t" :: "r" (uCR0));
1701# endif
1702# else
1703 __asm
1704 {
1705# ifdef RT_ARCH_AMD64
1706 mov rax, [uCR0]
1707 mov cr0, rax
1708# else
1709 mov eax, [uCR0]
1710 mov cr0, eax
1711# endif
1712 }
1713# endif
1714}
1715#endif
1716
1717
1718/**
1719 * Get cr2.
1720 * @returns cr2.
1721 */
1722#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1723RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTXREG) ASMGetCR2(void);
1724#else
1725DECLINLINE(RTCCUINTXREG) ASMGetCR2(void)
1726{
1727 RTCCUINTXREG uCR2;
1728# if RT_INLINE_ASM_USES_INTRIN
1729 uCR2 = __readcr2();
1730
1731# elif RT_INLINE_ASM_GNU_STYLE
1732# ifdef RT_ARCH_AMD64
1733 __asm__ __volatile__("movq %%cr2, %0\t\n" : "=r" (uCR2));
1734# else
1735 __asm__ __volatile__("movl %%cr2, %0\t\n" : "=r" (uCR2));
1736# endif
1737# else
1738 __asm
1739 {
1740# ifdef RT_ARCH_AMD64
1741 mov rax, cr2
1742 mov [uCR2], rax
1743# else
1744 mov eax, cr2
1745 mov [uCR2], eax
1746# endif
1747 }
1748# endif
1749 return uCR2;
1750}
1751#endif
1752
1753
1754/**
1755 * Sets the CR2 register.
1756 * @param uCR2 The new CR0 value.
1757 */
1758#if RT_INLINE_ASM_EXTERNAL
1759RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetCR2(RTCCUINTXREG uCR2);
1760#else
1761DECLINLINE(void) ASMSetCR2(RTCCUINTXREG uCR2)
1762{
1763# if RT_INLINE_ASM_GNU_STYLE
1764# ifdef RT_ARCH_AMD64
1765 __asm__ __volatile__("movq %0, %%cr2\n\t" :: "r" (uCR2));
1766# else
1767 __asm__ __volatile__("movl %0, %%cr2\n\t" :: "r" (uCR2));
1768# endif
1769# else
1770 __asm
1771 {
1772# ifdef RT_ARCH_AMD64
1773 mov rax, [uCR2]
1774 mov cr2, rax
1775# else
1776 mov eax, [uCR2]
1777 mov cr2, eax
1778# endif
1779 }
1780# endif
1781}
1782#endif
1783
1784
1785/**
1786 * Get cr3.
1787 * @returns cr3.
1788 */
1789#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1790RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTXREG) ASMGetCR3(void);
1791#else
1792DECLINLINE(RTCCUINTXREG) ASMGetCR3(void)
1793{
1794 RTCCUINTXREG uCR3;
1795# if RT_INLINE_ASM_USES_INTRIN
1796 uCR3 = __readcr3();
1797
1798# elif RT_INLINE_ASM_GNU_STYLE
1799# ifdef RT_ARCH_AMD64
1800 __asm__ __volatile__("movq %%cr3, %0\t\n" : "=r" (uCR3));
1801# else
1802 __asm__ __volatile__("movl %%cr3, %0\t\n" : "=r" (uCR3));
1803# endif
1804# else
1805 __asm
1806 {
1807# ifdef RT_ARCH_AMD64
1808 mov rax, cr3
1809 mov [uCR3], rax
1810# else
1811 mov eax, cr3
1812 mov [uCR3], eax
1813# endif
1814 }
1815# endif
1816 return uCR3;
1817}
1818#endif
1819
1820
1821/**
1822 * Sets the CR3 register.
1823 *
1824 * @param uCR3 New CR3 value.
1825 */
1826#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1827RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetCR3(RTCCUINTXREG uCR3);
1828#else
1829DECLINLINE(void) ASMSetCR3(RTCCUINTXREG uCR3)
1830{
1831# if RT_INLINE_ASM_USES_INTRIN
1832 __writecr3(uCR3);
1833
1834# elif RT_INLINE_ASM_GNU_STYLE
1835# ifdef RT_ARCH_AMD64
1836 __asm__ __volatile__("movq %0, %%cr3\n\t" : : "r" (uCR3));
1837# else
1838 __asm__ __volatile__("movl %0, %%cr3\n\t" : : "r" (uCR3));
1839# endif
1840# else
1841 __asm
1842 {
1843# ifdef RT_ARCH_AMD64
1844 mov rax, [uCR3]
1845 mov cr3, rax
1846# else
1847 mov eax, [uCR3]
1848 mov cr3, eax
1849# endif
1850 }
1851# endif
1852}
1853#endif
1854
1855
1856/**
1857 * Reloads the CR3 register.
1858 */
1859#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1860RT_ASM_DECL_PRAGMA_WATCOM(void) ASMReloadCR3(void);
1861#else
1862DECLINLINE(void) ASMReloadCR3(void)
1863{
1864# if RT_INLINE_ASM_USES_INTRIN
1865 __writecr3(__readcr3());
1866
1867# elif RT_INLINE_ASM_GNU_STYLE
1868 RTCCUINTXREG u;
1869# ifdef RT_ARCH_AMD64
1870 __asm__ __volatile__("movq %%cr3, %0\n\t"
1871 "movq %0, %%cr3\n\t"
1872 : "=r" (u));
1873# else
1874 __asm__ __volatile__("movl %%cr3, %0\n\t"
1875 "movl %0, %%cr3\n\t"
1876 : "=r" (u));
1877# endif
1878# else
1879 __asm
1880 {
1881# ifdef RT_ARCH_AMD64
1882 mov rax, cr3
1883 mov cr3, rax
1884# else
1885 mov eax, cr3
1886 mov cr3, eax
1887# endif
1888 }
1889# endif
1890}
1891#endif
1892
1893
1894/**
1895 * Get cr4.
1896 * @returns cr4.
1897 */
1898#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1899RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTXREG) ASMGetCR4(void);
1900#else
1901DECLINLINE(RTCCUINTXREG) ASMGetCR4(void)
1902{
1903 RTCCUINTXREG uCR4;
1904# if RT_INLINE_ASM_USES_INTRIN
1905 uCR4 = __readcr4();
1906
1907# elif RT_INLINE_ASM_GNU_STYLE
1908# ifdef RT_ARCH_AMD64
1909 __asm__ __volatile__("movq %%cr4, %0\t\n" : "=r" (uCR4));
1910# else
1911 __asm__ __volatile__("movl %%cr4, %0\t\n" : "=r" (uCR4));
1912# endif
1913# else
1914 __asm
1915 {
1916# ifdef RT_ARCH_AMD64
1917 mov rax, cr4
1918 mov [uCR4], rax
1919# else
1920 push eax /* just in case */
1921 /*mov eax, cr4*/
1922 _emit 0x0f
1923 _emit 0x20
1924 _emit 0xe0
1925 mov [uCR4], eax
1926 pop eax
1927# endif
1928 }
1929# endif
1930 return uCR4;
1931}
1932#endif
1933
1934
1935/**
1936 * Sets the CR4 register.
1937 *
1938 * @param uCR4 New CR4 value.
1939 */
1940#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1941RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetCR4(RTCCUINTXREG uCR4);
1942#else
1943DECLINLINE(void) ASMSetCR4(RTCCUINTXREG uCR4)
1944{
1945# if RT_INLINE_ASM_USES_INTRIN
1946 __writecr4(uCR4);
1947
1948# elif RT_INLINE_ASM_GNU_STYLE
1949# ifdef RT_ARCH_AMD64
1950 __asm__ __volatile__("movq %0, %%cr4\n\t" : : "r" (uCR4));
1951# else
1952 __asm__ __volatile__("movl %0, %%cr4\n\t" : : "r" (uCR4));
1953# endif
1954# else
1955 __asm
1956 {
1957# ifdef RT_ARCH_AMD64
1958 mov rax, [uCR4]
1959 mov cr4, rax
1960# else
1961 mov eax, [uCR4]
1962 _emit 0x0F
1963 _emit 0x22
1964 _emit 0xE0 /* mov cr4, eax */
1965# endif
1966 }
1967# endif
1968}
1969#endif
1970
1971
1972/**
1973 * Get cr8.
1974 * @returns cr8.
1975 * @remark The lock prefix hack for access from non-64-bit modes is NOT used and 0 is returned.
1976 */
1977#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
1978DECLASM(RTCCUINTXREG) ASMGetCR8(void);
1979#else
1980DECLINLINE(RTCCUINTXREG) ASMGetCR8(void)
1981{
1982# ifdef RT_ARCH_AMD64
1983 RTCCUINTXREG uCR8;
1984# if RT_INLINE_ASM_USES_INTRIN
1985 uCR8 = __readcr8();
1986
1987# elif RT_INLINE_ASM_GNU_STYLE
1988 __asm__ __volatile__("movq %%cr8, %0\t\n" : "=r" (uCR8));
1989# else
1990 __asm
1991 {
1992 mov rax, cr8
1993 mov [uCR8], rax
1994 }
1995# endif
1996 return uCR8;
1997# else /* !RT_ARCH_AMD64 */
1998 return 0;
1999# endif /* !RT_ARCH_AMD64 */
2000}
2001#endif
2002
2003
2004/**
2005 * Get XCR0 (eXtended feature Control Register 0).
2006 * @returns xcr0.
2007 */
2008DECLASM(uint64_t) ASMGetXcr0(void);
2009
2010/**
2011 * Sets the XCR0 register.
2012 * @param uXcr0 The new XCR0 value.
2013 */
2014DECLASM(void) ASMSetXcr0(uint64_t uXcr0);
2015
2016struct X86XSAVEAREA;
2017/**
2018 * Save extended CPU state.
2019 * @param pXStateArea Where to save the state.
2020 * @param fComponents Which state components to save.
2021 */
2022DECLASM(void) ASMXSave(struct X86XSAVEAREA RT_FAR *pXStateArea, uint64_t fComponents);
2023
2024/**
2025 * Loads extended CPU state.
2026 * @param pXStateArea Where to load the state from.
2027 * @param fComponents Which state components to load.
2028 */
2029DECLASM(void) ASMXRstor(struct X86XSAVEAREA const RT_FAR *pXStateArea, uint64_t fComponents);
2030
2031
2032struct X86FXSTATE;
2033/**
2034 * Save FPU and SSE CPU state.
2035 * @param pXStateArea Where to save the state.
2036 */
2037DECLASM(void) ASMFxSave(struct X86FXSTATE RT_FAR *pXStateArea);
2038
2039/**
2040 * Load FPU and SSE CPU state.
2041 * @param pXStateArea Where to load the state from.
2042 */
2043DECLASM(void) ASMFxRstor(struct X86FXSTATE const RT_FAR *pXStateArea);
2044
2045
2046/**
2047 * Enables interrupts (EFLAGS.IF).
2048 */
2049#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2050RT_ASM_DECL_PRAGMA_WATCOM(void) ASMIntEnable(void);
2051#else
2052DECLINLINE(void) ASMIntEnable(void)
2053{
2054# if RT_INLINE_ASM_GNU_STYLE
2055 __asm("sti\n");
2056# elif RT_INLINE_ASM_USES_INTRIN
2057 _enable();
2058# else
2059 __asm sti
2060# endif
2061}
2062#endif
2063
2064
2065/**
2066 * Disables interrupts (!EFLAGS.IF).
2067 */
2068#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2069RT_ASM_DECL_PRAGMA_WATCOM(void) ASMIntDisable(void);
2070#else
2071DECLINLINE(void) ASMIntDisable(void)
2072{
2073# if RT_INLINE_ASM_GNU_STYLE
2074 __asm("cli\n");
2075# elif RT_INLINE_ASM_USES_INTRIN
2076 _disable();
2077# else
2078 __asm cli
2079# endif
2080}
2081#endif
2082
2083
2084/**
2085 * Disables interrupts and returns previous xFLAGS.
2086 */
2087#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2088RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTREG) ASMIntDisableFlags(void);
2089#else
2090DECLINLINE(RTCCUINTREG) ASMIntDisableFlags(void)
2091{
2092 RTCCUINTREG xFlags;
2093# if RT_INLINE_ASM_GNU_STYLE
2094# ifdef RT_ARCH_AMD64
2095 __asm__ __volatile__("pushfq\n\t"
2096 "cli\n\t"
2097 "popq %0\n\t"
2098 : "=r" (xFlags));
2099# else
2100 __asm__ __volatile__("pushfl\n\t"
2101 "cli\n\t"
2102 "popl %0\n\t"
2103 : "=r" (xFlags));
2104# endif
2105# elif RT_INLINE_ASM_USES_INTRIN && !defined(RT_ARCH_X86)
2106 xFlags = ASMGetFlags();
2107 _disable();
2108# else
2109 __asm {
2110 pushfd
2111 cli
2112 pop [xFlags]
2113 }
2114# endif
2115 return xFlags;
2116}
2117#endif
2118
2119
2120/**
2121 * Are interrupts enabled?
2122 *
2123 * @returns true / false.
2124 */
2125DECLINLINE(bool) ASMIntAreEnabled(void)
2126{
2127 RTCCUINTREG uFlags = ASMGetFlags();
2128 return uFlags & 0x200 /* X86_EFL_IF */ ? true : false;
2129}
2130
2131
2132/**
2133 * Halts the CPU until interrupted.
2134 */
2135#if RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2005
2136RT_ASM_DECL_PRAGMA_WATCOM(void) ASMHalt(void);
2137#else
2138DECLINLINE(void) ASMHalt(void)
2139{
2140# if RT_INLINE_ASM_GNU_STYLE
2141 __asm__ __volatile__("hlt\n\t");
2142# elif RT_INLINE_ASM_USES_INTRIN
2143 __halt();
2144# else
2145 __asm {
2146 hlt
2147 }
2148# endif
2149}
2150#endif
2151
2152
2153/**
2154 * Reads a machine specific register.
2155 *
2156 * @returns Register content.
2157 * @param uRegister Register to read.
2158 */
2159#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2160RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMRdMsr(uint32_t uRegister);
2161#else
2162DECLINLINE(uint64_t) ASMRdMsr(uint32_t uRegister)
2163{
2164 RTUINT64U u;
2165# if RT_INLINE_ASM_GNU_STYLE
2166 __asm__ __volatile__("rdmsr\n\t"
2167 : "=a" (u.s.Lo),
2168 "=d" (u.s.Hi)
2169 : "c" (uRegister));
2170
2171# elif RT_INLINE_ASM_USES_INTRIN
2172 u.u = __readmsr(uRegister);
2173
2174# else
2175 __asm
2176 {
2177 mov ecx, [uRegister]
2178 rdmsr
2179 mov [u.s.Lo], eax
2180 mov [u.s.Hi], edx
2181 }
2182# endif
2183
2184 return u.u;
2185}
2186#endif
2187
2188
2189/**
2190 * Writes a machine specific register.
2191 *
2192 * @param uRegister Register to write to.
2193 * @param u64Val Value to write.
2194 */
2195#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2196RT_ASM_DECL_PRAGMA_WATCOM_386(void) ASMWrMsr(uint32_t uRegister, uint64_t u64Val);
2197#else
2198DECLINLINE(void) ASMWrMsr(uint32_t uRegister, uint64_t u64Val)
2199{
2200 RTUINT64U u;
2201
2202 u.u = u64Val;
2203# if RT_INLINE_ASM_GNU_STYLE
2204 __asm__ __volatile__("wrmsr\n\t"
2205 ::"a" (u.s.Lo),
2206 "d" (u.s.Hi),
2207 "c" (uRegister));
2208
2209# elif RT_INLINE_ASM_USES_INTRIN
2210 __writemsr(uRegister, u.u);
2211
2212# else
2213 __asm
2214 {
2215 mov ecx, [uRegister]
2216 mov edx, [u.s.Hi]
2217 mov eax, [u.s.Lo]
2218 wrmsr
2219 }
2220# endif
2221}
2222#endif
2223
2224
2225/**
2226 * Reads a machine specific register, extended version (for AMD).
2227 *
2228 * @returns Register content.
2229 * @param uRegister Register to read.
2230 * @param uXDI RDI/EDI value.
2231 */
2232#if RT_INLINE_ASM_EXTERNAL
2233RT_ASM_DECL_PRAGMA_WATCOM_386(uint64_t) ASMRdMsrEx(uint32_t uRegister, RTCCUINTXREG uXDI);
2234#else
2235DECLINLINE(uint64_t) ASMRdMsrEx(uint32_t uRegister, RTCCUINTXREG uXDI)
2236{
2237 RTUINT64U u;
2238# if RT_INLINE_ASM_GNU_STYLE
2239 __asm__ __volatile__("rdmsr\n\t"
2240 : "=a" (u.s.Lo),
2241 "=d" (u.s.Hi)
2242 : "c" (uRegister),
2243 "D" (uXDI));
2244
2245# else
2246 __asm
2247 {
2248 mov ecx, [uRegister]
2249 xchg edi, [uXDI]
2250 rdmsr
2251 mov [u.s.Lo], eax
2252 mov [u.s.Hi], edx
2253 xchg edi, [uXDI]
2254 }
2255# endif
2256
2257 return u.u;
2258}
2259#endif
2260
2261
2262/**
2263 * Writes a machine specific register, extended version (for AMD).
2264 *
2265 * @param uRegister Register to write to.
2266 * @param uXDI RDI/EDI value.
2267 * @param u64Val Value to write.
2268 */
2269#if RT_INLINE_ASM_EXTERNAL
2270RT_ASM_DECL_PRAGMA_WATCOM_386(void) ASMWrMsrEx(uint32_t uRegister, RTCCUINTXREG uXDI, uint64_t u64Val);
2271#else
2272DECLINLINE(void) ASMWrMsrEx(uint32_t uRegister, RTCCUINTXREG uXDI, uint64_t u64Val)
2273{
2274 RTUINT64U u;
2275
2276 u.u = u64Val;
2277# if RT_INLINE_ASM_GNU_STYLE
2278 __asm__ __volatile__("wrmsr\n\t"
2279 ::"a" (u.s.Lo),
2280 "d" (u.s.Hi),
2281 "c" (uRegister),
2282 "D" (uXDI));
2283
2284# else
2285 __asm
2286 {
2287 mov ecx, [uRegister]
2288 xchg edi, [uXDI]
2289 mov edx, [u.s.Hi]
2290 mov eax, [u.s.Lo]
2291 wrmsr
2292 xchg edi, [uXDI]
2293 }
2294# endif
2295}
2296#endif
2297
2298
2299
2300/**
2301 * Reads low part of a machine specific register.
2302 *
2303 * @returns Register content.
2304 * @param uRegister Register to read.
2305 */
2306#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2307RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMRdMsr_Low(uint32_t uRegister);
2308#else
2309DECLINLINE(uint32_t) ASMRdMsr_Low(uint32_t uRegister)
2310{
2311 uint32_t u32;
2312# if RT_INLINE_ASM_GNU_STYLE
2313 __asm__ __volatile__("rdmsr\n\t"
2314 : "=a" (u32)
2315 : "c" (uRegister)
2316 : "edx");
2317
2318# elif RT_INLINE_ASM_USES_INTRIN
2319 u32 = (uint32_t)__readmsr(uRegister);
2320
2321#else
2322 __asm
2323 {
2324 mov ecx, [uRegister]
2325 rdmsr
2326 mov [u32], eax
2327 }
2328# endif
2329
2330 return u32;
2331}
2332#endif
2333
2334
2335/**
2336 * Reads high part of a machine specific register.
2337 *
2338 * @returns Register content.
2339 * @param uRegister Register to read.
2340 */
2341#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2342RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMRdMsr_High(uint32_t uRegister);
2343#else
2344DECLINLINE(uint32_t) ASMRdMsr_High(uint32_t uRegister)
2345{
2346 uint32_t u32;
2347# if RT_INLINE_ASM_GNU_STYLE
2348 __asm__ __volatile__("rdmsr\n\t"
2349 : "=d" (u32)
2350 : "c" (uRegister)
2351 : "eax");
2352
2353# elif RT_INLINE_ASM_USES_INTRIN
2354 u32 = (uint32_t)(__readmsr(uRegister) >> 32);
2355
2356# else
2357 __asm
2358 {
2359 mov ecx, [uRegister]
2360 rdmsr
2361 mov [u32], edx
2362 }
2363# endif
2364
2365 return u32;
2366}
2367#endif
2368
2369
2370/**
2371 * Gets dr0.
2372 *
2373 * @returns dr0.
2374 */
2375#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2376RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTXREG) ASMGetDR0(void);
2377#else
2378DECLINLINE(RTCCUINTXREG) ASMGetDR0(void)
2379{
2380 RTCCUINTXREG uDR0;
2381# if RT_INLINE_ASM_USES_INTRIN
2382 uDR0 = __readdr(0);
2383# elif RT_INLINE_ASM_GNU_STYLE
2384# ifdef RT_ARCH_AMD64
2385 __asm__ __volatile__("movq %%dr0, %0\n\t" : "=r" (uDR0));
2386# else
2387 __asm__ __volatile__("movl %%dr0, %0\n\t" : "=r" (uDR0));
2388# endif
2389# else
2390 __asm
2391 {
2392# ifdef RT_ARCH_AMD64
2393 mov rax, dr0
2394 mov [uDR0], rax
2395# else
2396 mov eax, dr0
2397 mov [uDR0], eax
2398# endif
2399 }
2400# endif
2401 return uDR0;
2402}
2403#endif
2404
2405
2406/**
2407 * Gets dr1.
2408 *
2409 * @returns dr1.
2410 */
2411#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2412RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTXREG) ASMGetDR1(void);
2413#else
2414DECLINLINE(RTCCUINTXREG) ASMGetDR1(void)
2415{
2416 RTCCUINTXREG uDR1;
2417# if RT_INLINE_ASM_USES_INTRIN
2418 uDR1 = __readdr(1);
2419# elif RT_INLINE_ASM_GNU_STYLE
2420# ifdef RT_ARCH_AMD64
2421 __asm__ __volatile__("movq %%dr1, %0\n\t" : "=r" (uDR1));
2422# else
2423 __asm__ __volatile__("movl %%dr1, %0\n\t" : "=r" (uDR1));
2424# endif
2425# else
2426 __asm
2427 {
2428# ifdef RT_ARCH_AMD64
2429 mov rax, dr1
2430 mov [uDR1], rax
2431# else
2432 mov eax, dr1
2433 mov [uDR1], eax
2434# endif
2435 }
2436# endif
2437 return uDR1;
2438}
2439#endif
2440
2441
2442/**
2443 * Gets dr2.
2444 *
2445 * @returns dr2.
2446 */
2447#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2448RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTXREG) ASMGetDR2(void);
2449#else
2450DECLINLINE(RTCCUINTXREG) ASMGetDR2(void)
2451{
2452 RTCCUINTXREG uDR2;
2453# if RT_INLINE_ASM_USES_INTRIN
2454 uDR2 = __readdr(2);
2455# elif RT_INLINE_ASM_GNU_STYLE
2456# ifdef RT_ARCH_AMD64
2457 __asm__ __volatile__("movq %%dr2, %0\n\t" : "=r" (uDR2));
2458# else
2459 __asm__ __volatile__("movl %%dr2, %0\n\t" : "=r" (uDR2));
2460# endif
2461# else
2462 __asm
2463 {
2464# ifdef RT_ARCH_AMD64
2465 mov rax, dr2
2466 mov [uDR2], rax
2467# else
2468 mov eax, dr2
2469 mov [uDR2], eax
2470# endif
2471 }
2472# endif
2473 return uDR2;
2474}
2475#endif
2476
2477
2478/**
2479 * Gets dr3.
2480 *
2481 * @returns dr3.
2482 */
2483#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2484RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTXREG) ASMGetDR3(void);
2485#else
2486DECLINLINE(RTCCUINTXREG) ASMGetDR3(void)
2487{
2488 RTCCUINTXREG uDR3;
2489# if RT_INLINE_ASM_USES_INTRIN
2490 uDR3 = __readdr(3);
2491# elif RT_INLINE_ASM_GNU_STYLE
2492# ifdef RT_ARCH_AMD64
2493 __asm__ __volatile__("movq %%dr3, %0\n\t" : "=r" (uDR3));
2494# else
2495 __asm__ __volatile__("movl %%dr3, %0\n\t" : "=r" (uDR3));
2496# endif
2497# else
2498 __asm
2499 {
2500# ifdef RT_ARCH_AMD64
2501 mov rax, dr3
2502 mov [uDR3], rax
2503# else
2504 mov eax, dr3
2505 mov [uDR3], eax
2506# endif
2507 }
2508# endif
2509 return uDR3;
2510}
2511#endif
2512
2513
2514/**
2515 * Gets dr6.
2516 *
2517 * @returns dr6.
2518 */
2519#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2520RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTXREG) ASMGetDR6(void);
2521#else
2522DECLINLINE(RTCCUINTXREG) ASMGetDR6(void)
2523{
2524 RTCCUINTXREG uDR6;
2525# if RT_INLINE_ASM_USES_INTRIN
2526 uDR6 = __readdr(6);
2527# elif RT_INLINE_ASM_GNU_STYLE
2528# ifdef RT_ARCH_AMD64
2529 __asm__ __volatile__("movq %%dr6, %0\n\t" : "=r" (uDR6));
2530# else
2531 __asm__ __volatile__("movl %%dr6, %0\n\t" : "=r" (uDR6));
2532# endif
2533# else
2534 __asm
2535 {
2536# ifdef RT_ARCH_AMD64
2537 mov rax, dr6
2538 mov [uDR6], rax
2539# else
2540 mov eax, dr6
2541 mov [uDR6], eax
2542# endif
2543 }
2544# endif
2545 return uDR6;
2546}
2547#endif
2548
2549
2550/**
2551 * Reads and clears DR6.
2552 *
2553 * @returns DR6.
2554 */
2555#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2556RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTXREG) ASMGetAndClearDR6(void);
2557#else
2558DECLINLINE(RTCCUINTXREG) ASMGetAndClearDR6(void)
2559{
2560 RTCCUINTXREG uDR6;
2561# if RT_INLINE_ASM_USES_INTRIN
2562 uDR6 = __readdr(6);
2563 __writedr(6, 0xffff0ff0U); /* 31-16 and 4-11 are 1's, 12 and 63-31 are zero. */
2564# elif RT_INLINE_ASM_GNU_STYLE
2565 RTCCUINTXREG uNewValue = 0xffff0ff0U;/* 31-16 and 4-11 are 1's, 12 and 63-31 are zero. */
2566# ifdef RT_ARCH_AMD64
2567 __asm__ __volatile__("movq %%dr6, %0\n\t"
2568 "movq %1, %%dr6\n\t"
2569 : "=r" (uDR6)
2570 : "r" (uNewValue));
2571# else
2572 __asm__ __volatile__("movl %%dr6, %0\n\t"
2573 "movl %1, %%dr6\n\t"
2574 : "=r" (uDR6)
2575 : "r" (uNewValue));
2576# endif
2577# else
2578 __asm
2579 {
2580# ifdef RT_ARCH_AMD64
2581 mov rax, dr6
2582 mov [uDR6], rax
2583 mov rcx, rax
2584 mov ecx, 0ffff0ff0h; /* 31-16 and 4-11 are 1's, 12 and 63-31 are zero. */
2585 mov dr6, rcx
2586# else
2587 mov eax, dr6
2588 mov [uDR6], eax
2589 mov ecx, 0ffff0ff0h; /* 31-16 and 4-11 are 1's, 12 is zero. */
2590 mov dr6, ecx
2591# endif
2592 }
2593# endif
2594 return uDR6;
2595}
2596#endif
2597
2598
2599/**
2600 * Gets dr7.
2601 *
2602 * @returns dr7.
2603 */
2604#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2605RT_ASM_DECL_PRAGMA_WATCOM(RTCCUINTXREG) ASMGetDR7(void);
2606#else
2607DECLINLINE(RTCCUINTXREG) ASMGetDR7(void)
2608{
2609 RTCCUINTXREG uDR7;
2610# if RT_INLINE_ASM_USES_INTRIN
2611 uDR7 = __readdr(7);
2612# elif RT_INLINE_ASM_GNU_STYLE
2613# ifdef RT_ARCH_AMD64
2614 __asm__ __volatile__("movq %%dr7, %0\n\t" : "=r" (uDR7));
2615# else
2616 __asm__ __volatile__("movl %%dr7, %0\n\t" : "=r" (uDR7));
2617# endif
2618# else
2619 __asm
2620 {
2621# ifdef RT_ARCH_AMD64
2622 mov rax, dr7
2623 mov [uDR7], rax
2624# else
2625 mov eax, dr7
2626 mov [uDR7], eax
2627# endif
2628 }
2629# endif
2630 return uDR7;
2631}
2632#endif
2633
2634
2635/**
2636 * Sets dr0.
2637 *
2638 * @param uDRVal Debug register value to write
2639 */
2640#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2641RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetDR0(RTCCUINTXREG uDRVal);
2642#else
2643DECLINLINE(void) ASMSetDR0(RTCCUINTXREG uDRVal)
2644{
2645# if RT_INLINE_ASM_USES_INTRIN
2646 __writedr(0, uDRVal);
2647# elif RT_INLINE_ASM_GNU_STYLE
2648# ifdef RT_ARCH_AMD64
2649 __asm__ __volatile__("movq %0, %%dr0\n\t" : : "r" (uDRVal));
2650# else
2651 __asm__ __volatile__("movl %0, %%dr0\n\t" : : "r" (uDRVal));
2652# endif
2653# else
2654 __asm
2655 {
2656# ifdef RT_ARCH_AMD64
2657 mov rax, [uDRVal]
2658 mov dr0, rax
2659# else
2660 mov eax, [uDRVal]
2661 mov dr0, eax
2662# endif
2663 }
2664# endif
2665}
2666#endif
2667
2668
2669/**
2670 * Sets dr1.
2671 *
2672 * @param uDRVal Debug register value to write
2673 */
2674#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2675RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetDR1(RTCCUINTXREG uDRVal);
2676#else
2677DECLINLINE(void) ASMSetDR1(RTCCUINTXREG uDRVal)
2678{
2679# if RT_INLINE_ASM_USES_INTRIN
2680 __writedr(1, uDRVal);
2681# elif RT_INLINE_ASM_GNU_STYLE
2682# ifdef RT_ARCH_AMD64
2683 __asm__ __volatile__("movq %0, %%dr1\n\t" : : "r" (uDRVal));
2684# else
2685 __asm__ __volatile__("movl %0, %%dr1\n\t" : : "r" (uDRVal));
2686# endif
2687# else
2688 __asm
2689 {
2690# ifdef RT_ARCH_AMD64
2691 mov rax, [uDRVal]
2692 mov dr1, rax
2693# else
2694 mov eax, [uDRVal]
2695 mov dr1, eax
2696# endif
2697 }
2698# endif
2699}
2700#endif
2701
2702
2703/**
2704 * Sets dr2.
2705 *
2706 * @param uDRVal Debug register value to write
2707 */
2708#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2709RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetDR2(RTCCUINTXREG uDRVal);
2710#else
2711DECLINLINE(void) ASMSetDR2(RTCCUINTXREG uDRVal)
2712{
2713# if RT_INLINE_ASM_USES_INTRIN
2714 __writedr(2, uDRVal);
2715# elif RT_INLINE_ASM_GNU_STYLE
2716# ifdef RT_ARCH_AMD64
2717 __asm__ __volatile__("movq %0, %%dr2\n\t" : : "r" (uDRVal));
2718# else
2719 __asm__ __volatile__("movl %0, %%dr2\n\t" : : "r" (uDRVal));
2720# endif
2721# else
2722 __asm
2723 {
2724# ifdef RT_ARCH_AMD64
2725 mov rax, [uDRVal]
2726 mov dr2, rax
2727# else
2728 mov eax, [uDRVal]
2729 mov dr2, eax
2730# endif
2731 }
2732# endif
2733}
2734#endif
2735
2736
2737/**
2738 * Sets dr3.
2739 *
2740 * @param uDRVal Debug register value to write
2741 */
2742#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2743RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetDR3(RTCCUINTXREG uDRVal);
2744#else
2745DECLINLINE(void) ASMSetDR3(RTCCUINTXREG uDRVal)
2746{
2747# if RT_INLINE_ASM_USES_INTRIN
2748 __writedr(3, uDRVal);
2749# elif RT_INLINE_ASM_GNU_STYLE
2750# ifdef RT_ARCH_AMD64
2751 __asm__ __volatile__("movq %0, %%dr3\n\t" : : "r" (uDRVal));
2752# else
2753 __asm__ __volatile__("movl %0, %%dr3\n\t" : : "r" (uDRVal));
2754# endif
2755# else
2756 __asm
2757 {
2758# ifdef RT_ARCH_AMD64
2759 mov rax, [uDRVal]
2760 mov dr3, rax
2761# else
2762 mov eax, [uDRVal]
2763 mov dr3, eax
2764# endif
2765 }
2766# endif
2767}
2768#endif
2769
2770
2771/**
2772 * Sets dr6.
2773 *
2774 * @param uDRVal Debug register value to write
2775 */
2776#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2777RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetDR6(RTCCUINTXREG uDRVal);
2778#else
2779DECLINLINE(void) ASMSetDR6(RTCCUINTXREG uDRVal)
2780{
2781# if RT_INLINE_ASM_USES_INTRIN
2782 __writedr(6, uDRVal);
2783# elif RT_INLINE_ASM_GNU_STYLE
2784# ifdef RT_ARCH_AMD64
2785 __asm__ __volatile__("movq %0, %%dr6\n\t" : : "r" (uDRVal));
2786# else
2787 __asm__ __volatile__("movl %0, %%dr6\n\t" : : "r" (uDRVal));
2788# endif
2789# else
2790 __asm
2791 {
2792# ifdef RT_ARCH_AMD64
2793 mov rax, [uDRVal]
2794 mov dr6, rax
2795# else
2796 mov eax, [uDRVal]
2797 mov dr6, eax
2798# endif
2799 }
2800# endif
2801}
2802#endif
2803
2804
2805/**
2806 * Sets dr7.
2807 *
2808 * @param uDRVal Debug register value to write
2809 */
2810#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2811RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSetDR7(RTCCUINTXREG uDRVal);
2812#else
2813DECLINLINE(void) ASMSetDR7(RTCCUINTXREG uDRVal)
2814{
2815# if RT_INLINE_ASM_USES_INTRIN
2816 __writedr(7, uDRVal);
2817# elif RT_INLINE_ASM_GNU_STYLE
2818# ifdef RT_ARCH_AMD64
2819 __asm__ __volatile__("movq %0, %%dr7\n\t" : : "r" (uDRVal));
2820# else
2821 __asm__ __volatile__("movl %0, %%dr7\n\t" : : "r" (uDRVal));
2822# endif
2823# else
2824 __asm
2825 {
2826# ifdef RT_ARCH_AMD64
2827 mov rax, [uDRVal]
2828 mov dr7, rax
2829# else
2830 mov eax, [uDRVal]
2831 mov dr7, eax
2832# endif
2833 }
2834# endif
2835}
2836#endif
2837
2838
2839/**
2840 * Writes a 8-bit unsigned integer to an I/O port, ordered.
2841 *
2842 * @param Port I/O port to write to.
2843 * @param u8 8-bit integer to write.
2844 */
2845#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2846RT_ASM_DECL_PRAGMA_WATCOM(void) ASMOutU8(RTIOPORT Port, uint8_t u8);
2847#else
2848DECLINLINE(void) ASMOutU8(RTIOPORT Port, uint8_t u8)
2849{
2850# if RT_INLINE_ASM_GNU_STYLE
2851 __asm__ __volatile__("outb %b1, %w0\n\t"
2852 :: "Nd" (Port),
2853 "a" (u8));
2854
2855# elif RT_INLINE_ASM_USES_INTRIN
2856 __outbyte(Port, u8);
2857
2858# else
2859 __asm
2860 {
2861 mov dx, [Port]
2862 mov al, [u8]
2863 out dx, al
2864 }
2865# endif
2866}
2867#endif
2868
2869
2870/**
2871 * Reads a 8-bit unsigned integer from an I/O port, ordered.
2872 *
2873 * @returns 8-bit integer.
2874 * @param Port I/O port to read from.
2875 */
2876#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2877RT_ASM_DECL_PRAGMA_WATCOM(uint8_t) ASMInU8(RTIOPORT Port);
2878#else
2879DECLINLINE(uint8_t) ASMInU8(RTIOPORT Port)
2880{
2881 uint8_t u8;
2882# if RT_INLINE_ASM_GNU_STYLE
2883 __asm__ __volatile__("inb %w1, %b0\n\t"
2884 : "=a" (u8)
2885 : "Nd" (Port));
2886
2887# elif RT_INLINE_ASM_USES_INTRIN
2888 u8 = __inbyte(Port);
2889
2890# else
2891 __asm
2892 {
2893 mov dx, [Port]
2894 in al, dx
2895 mov [u8], al
2896 }
2897# endif
2898 return u8;
2899}
2900#endif
2901
2902
2903/**
2904 * Writes a 16-bit unsigned integer to an I/O port, ordered.
2905 *
2906 * @param Port I/O port to write to.
2907 * @param u16 16-bit integer to write.
2908 */
2909#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2910RT_ASM_DECL_PRAGMA_WATCOM(void) ASMOutU16(RTIOPORT Port, uint16_t u16);
2911#else
2912DECLINLINE(void) ASMOutU16(RTIOPORT Port, uint16_t u16)
2913{
2914# if RT_INLINE_ASM_GNU_STYLE
2915 __asm__ __volatile__("outw %w1, %w0\n\t"
2916 :: "Nd" (Port),
2917 "a" (u16));
2918
2919# elif RT_INLINE_ASM_USES_INTRIN
2920 __outword(Port, u16);
2921
2922# else
2923 __asm
2924 {
2925 mov dx, [Port]
2926 mov ax, [u16]
2927 out dx, ax
2928 }
2929# endif
2930}
2931#endif
2932
2933
2934/**
2935 * Reads a 16-bit unsigned integer from an I/O port, ordered.
2936 *
2937 * @returns 16-bit integer.
2938 * @param Port I/O port to read from.
2939 */
2940#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2941RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMInU16(RTIOPORT Port);
2942#else
2943DECLINLINE(uint16_t) ASMInU16(RTIOPORT Port)
2944{
2945 uint16_t u16;
2946# if RT_INLINE_ASM_GNU_STYLE
2947 __asm__ __volatile__("inw %w1, %w0\n\t"
2948 : "=a" (u16)
2949 : "Nd" (Port));
2950
2951# elif RT_INLINE_ASM_USES_INTRIN
2952 u16 = __inword(Port);
2953
2954# else
2955 __asm
2956 {
2957 mov dx, [Port]
2958 in ax, dx
2959 mov [u16], ax
2960 }
2961# endif
2962 return u16;
2963}
2964#endif
2965
2966
2967/**
2968 * Writes a 32-bit unsigned integer to an I/O port, ordered.
2969 *
2970 * @param Port I/O port to write to.
2971 * @param u32 32-bit integer to write.
2972 */
2973#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
2974RT_ASM_DECL_PRAGMA_WATCOM(void) ASMOutU32(RTIOPORT Port, uint32_t u32);
2975#else
2976DECLINLINE(void) ASMOutU32(RTIOPORT Port, uint32_t u32)
2977{
2978# if RT_INLINE_ASM_GNU_STYLE
2979 __asm__ __volatile__("outl %1, %w0\n\t"
2980 :: "Nd" (Port),
2981 "a" (u32));
2982
2983# elif RT_INLINE_ASM_USES_INTRIN
2984 __outdword(Port, u32);
2985
2986# else
2987 __asm
2988 {
2989 mov dx, [Port]
2990 mov eax, [u32]
2991 out dx, eax
2992 }
2993# endif
2994}
2995#endif
2996
2997
2998/**
2999 * Reads a 32-bit unsigned integer from an I/O port, ordered.
3000 *
3001 * @returns 32-bit integer.
3002 * @param Port I/O port to read from.
3003 */
3004#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
3005RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMInU32(RTIOPORT Port);
3006#else
3007DECLINLINE(uint32_t) ASMInU32(RTIOPORT Port)
3008{
3009 uint32_t u32;
3010# if RT_INLINE_ASM_GNU_STYLE
3011 __asm__ __volatile__("inl %w1, %0\n\t"
3012 : "=a" (u32)
3013 : "Nd" (Port));
3014
3015# elif RT_INLINE_ASM_USES_INTRIN
3016 u32 = __indword(Port);
3017
3018# else
3019 __asm
3020 {
3021 mov dx, [Port]
3022 in eax, dx
3023 mov [u32], eax
3024 }
3025# endif
3026 return u32;
3027}
3028#endif
3029
3030
3031/**
3032 * Writes a string of 8-bit unsigned integer items to an I/O port, ordered.
3033 *
3034 * @param Port I/O port to write to.
3035 * @param pau8 Pointer to the string buffer.
3036 * @param c The number of items to write.
3037 */
3038#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
3039RT_ASM_DECL_PRAGMA_WATCOM(void) ASMOutStrU8(RTIOPORT Port, uint8_t const RT_FAR *pau8, size_t c);
3040#else
3041DECLINLINE(void) ASMOutStrU8(RTIOPORT Port, uint8_t const RT_FAR *pau8, size_t c)
3042{
3043# if RT_INLINE_ASM_GNU_STYLE
3044 __asm__ __volatile__("rep; outsb\n\t"
3045 : "+S" (pau8),
3046 "+c" (c)
3047 : "d" (Port));
3048
3049# elif RT_INLINE_ASM_USES_INTRIN
3050 __outbytestring(Port, (unsigned char RT_FAR *)pau8, (unsigned long)c);
3051
3052# else
3053 __asm
3054 {
3055 mov dx, [Port]
3056 mov ecx, [c]
3057 mov eax, [pau8]
3058 xchg esi, eax
3059 rep outsb
3060 xchg esi, eax
3061 }
3062# endif
3063}
3064#endif
3065
3066
3067/**
3068 * Reads a string of 8-bit unsigned integer items from an I/O port, ordered.
3069 *
3070 * @param Port I/O port to read from.
3071 * @param pau8 Pointer to the string buffer (output).
3072 * @param c The number of items to read.
3073 */
3074#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
3075RT_ASM_DECL_PRAGMA_WATCOM(void) ASMInStrU8(RTIOPORT Port, uint8_t RT_FAR *pau8, size_t c);
3076#else
3077DECLINLINE(void) ASMInStrU8(RTIOPORT Port, uint8_t RT_FAR *pau8, size_t c)
3078{
3079# if RT_INLINE_ASM_GNU_STYLE
3080 __asm__ __volatile__("rep; insb\n\t"
3081 : "+D" (pau8),
3082 "+c" (c)
3083 : "d" (Port));
3084
3085# elif RT_INLINE_ASM_USES_INTRIN
3086 __inbytestring(Port, pau8, (unsigned long)c);
3087
3088# else
3089 __asm
3090 {
3091 mov dx, [Port]
3092 mov ecx, [c]
3093 mov eax, [pau8]
3094 xchg edi, eax
3095 rep insb
3096 xchg edi, eax
3097 }
3098# endif
3099}
3100#endif
3101
3102
3103/**
3104 * Writes a string of 16-bit unsigned integer items to an I/O port, ordered.
3105 *
3106 * @param Port I/O port to write to.
3107 * @param pau16 Pointer to the string buffer.
3108 * @param c The number of items to write.
3109 */
3110#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
3111RT_ASM_DECL_PRAGMA_WATCOM(void) ASMOutStrU16(RTIOPORT Port, uint16_t const RT_FAR *pau16, size_t c);
3112#else
3113DECLINLINE(void) ASMOutStrU16(RTIOPORT Port, uint16_t const RT_FAR *pau16, size_t c)
3114{
3115# if RT_INLINE_ASM_GNU_STYLE
3116 __asm__ __volatile__("rep; outsw\n\t"
3117 : "+S" (pau16),
3118 "+c" (c)
3119 : "d" (Port));
3120
3121# elif RT_INLINE_ASM_USES_INTRIN
3122 __outwordstring(Port, (unsigned short RT_FAR *)pau16, (unsigned long)c);
3123
3124# else
3125 __asm
3126 {
3127 mov dx, [Port]
3128 mov ecx, [c]
3129 mov eax, [pau16]
3130 xchg esi, eax
3131 rep outsw
3132 xchg esi, eax
3133 }
3134# endif
3135}
3136#endif
3137
3138
3139/**
3140 * Reads a string of 16-bit unsigned integer items from an I/O port, ordered.
3141 *
3142 * @param Port I/O port to read from.
3143 * @param pau16 Pointer to the string buffer (output).
3144 * @param c The number of items to read.
3145 */
3146#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
3147RT_ASM_DECL_PRAGMA_WATCOM(void) ASMInStrU16(RTIOPORT Port, uint16_t RT_FAR *pau16, size_t c);
3148#else
3149DECLINLINE(void) ASMInStrU16(RTIOPORT Port, uint16_t RT_FAR *pau16, size_t c)
3150{
3151# if RT_INLINE_ASM_GNU_STYLE
3152 __asm__ __volatile__("rep; insw\n\t"
3153 : "+D" (pau16),
3154 "+c" (c)
3155 : "d" (Port));
3156
3157# elif RT_INLINE_ASM_USES_INTRIN
3158 __inwordstring(Port, pau16, (unsigned long)c);
3159
3160# else
3161 __asm
3162 {
3163 mov dx, [Port]
3164 mov ecx, [c]
3165 mov eax, [pau16]
3166 xchg edi, eax
3167 rep insw
3168 xchg edi, eax
3169 }
3170# endif
3171}
3172#endif
3173
3174
3175/**
3176 * Writes a string of 32-bit unsigned integer items to an I/O port, ordered.
3177 *
3178 * @param Port I/O port to write to.
3179 * @param pau32 Pointer to the string buffer.
3180 * @param c The number of items to write.
3181 */
3182#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
3183RT_ASM_DECL_PRAGMA_WATCOM(void) ASMOutStrU32(RTIOPORT Port, uint32_t const RT_FAR *pau32, size_t c);
3184#else
3185DECLINLINE(void) ASMOutStrU32(RTIOPORT Port, uint32_t const RT_FAR *pau32, size_t c)
3186{
3187# if RT_INLINE_ASM_GNU_STYLE
3188 __asm__ __volatile__("rep; outsl\n\t"
3189 : "+S" (pau32),
3190 "+c" (c)
3191 : "d" (Port));
3192
3193# elif RT_INLINE_ASM_USES_INTRIN
3194 __outdwordstring(Port, (unsigned long RT_FAR *)pau32, (unsigned long)c);
3195
3196# else
3197 __asm
3198 {
3199 mov dx, [Port]
3200 mov ecx, [c]
3201 mov eax, [pau32]
3202 xchg esi, eax
3203 rep outsd
3204 xchg esi, eax
3205 }
3206# endif
3207}
3208#endif
3209
3210
3211/**
3212 * Reads a string of 32-bit unsigned integer items from an I/O port, ordered.
3213 *
3214 * @param Port I/O port to read from.
3215 * @param pau32 Pointer to the string buffer (output).
3216 * @param c The number of items to read.
3217 */
3218#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
3219RT_ASM_DECL_PRAGMA_WATCOM(void) ASMInStrU32(RTIOPORT Port, uint32_t RT_FAR *pau32, size_t c);
3220#else
3221DECLINLINE(void) ASMInStrU32(RTIOPORT Port, uint32_t RT_FAR *pau32, size_t c)
3222{
3223# if RT_INLINE_ASM_GNU_STYLE
3224 __asm__ __volatile__("rep; insl\n\t"
3225 : "+D" (pau32),
3226 "+c" (c)
3227 : "d" (Port));
3228
3229# elif RT_INLINE_ASM_USES_INTRIN
3230 __indwordstring(Port, (unsigned long RT_FAR *)pau32, (unsigned long)c);
3231
3232# else
3233 __asm
3234 {
3235 mov dx, [Port]
3236 mov ecx, [c]
3237 mov eax, [pau32]
3238 xchg edi, eax
3239 rep insd
3240 xchg edi, eax
3241 }
3242# endif
3243}
3244#endif
3245
3246
3247/**
3248 * Invalidate page.
3249 *
3250 * @param uPtr Address of the page to invalidate.
3251 */
3252#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
3253RT_ASM_DECL_PRAGMA_WATCOM(void) ASMInvalidatePage(RTCCUINTXREG uPtr);
3254#else
3255DECLINLINE(void) ASMInvalidatePage(RTCCUINTXREG uPtr)
3256{
3257# if RT_INLINE_ASM_USES_INTRIN
3258 __invlpg((void RT_FAR *)uPtr);
3259
3260# elif RT_INLINE_ASM_GNU_STYLE
3261 __asm__ __volatile__("invlpg %0\n\t"
3262 : : "m" (*(uint8_t RT_FAR *)(uintptr_t)uPtr));
3263# else
3264 __asm
3265 {
3266# ifdef RT_ARCH_AMD64
3267 mov rax, [uPtr]
3268 invlpg [rax]
3269# else
3270 mov eax, [uPtr]
3271 invlpg [eax]
3272# endif
3273 }
3274# endif
3275}
3276#endif
3277
3278
3279/**
3280 * Write back the internal caches and invalidate them.
3281 */
3282#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN
3283RT_ASM_DECL_PRAGMA_WATCOM(void) ASMWriteBackAndInvalidateCaches(void);
3284#else
3285DECLINLINE(void) ASMWriteBackAndInvalidateCaches(void)
3286{
3287# if RT_INLINE_ASM_USES_INTRIN
3288 __wbinvd();
3289
3290# elif RT_INLINE_ASM_GNU_STYLE
3291 __asm__ __volatile__("wbinvd");
3292# else
3293 __asm
3294 {
3295 wbinvd
3296 }
3297# endif
3298}
3299#endif
3300
3301
3302/**
3303 * Invalidate internal and (perhaps) external caches without first
3304 * flushing dirty cache lines. Use with extreme care.
3305 */
3306#if RT_INLINE_ASM_EXTERNAL
3307RT_ASM_DECL_PRAGMA_WATCOM(void) ASMInvalidateInternalCaches(void);
3308#else
3309DECLINLINE(void) ASMInvalidateInternalCaches(void)
3310{
3311# if RT_INLINE_ASM_GNU_STYLE
3312 __asm__ __volatile__("invd");
3313# else
3314 __asm
3315 {
3316 invd
3317 }
3318# endif
3319}
3320#endif
3321
3322
3323/**
3324 * Memory load/store fence, waits for any pending writes and reads to complete.
3325 * Requires the X86_CPUID_FEATURE_EDX_SSE2 CPUID bit set.
3326 */
3327DECLINLINE(void) ASMMemoryFenceSSE2(void)
3328{
3329#if RT_INLINE_ASM_GNU_STYLE
3330 __asm__ __volatile__ (".byte 0x0f,0xae,0xf0\n\t");
3331#elif RT_INLINE_ASM_USES_INTRIN
3332 _mm_mfence();
3333#else
3334 __asm
3335 {
3336 _emit 0x0f
3337 _emit 0xae
3338 _emit 0xf0
3339 }
3340#endif
3341}
3342
3343
3344/**
3345 * Memory store fence, waits for any writes to complete.
3346 * Requires the X86_CPUID_FEATURE_EDX_SSE CPUID bit set.
3347 */
3348DECLINLINE(void) ASMWriteFenceSSE(void)
3349{
3350#if RT_INLINE_ASM_GNU_STYLE
3351 __asm__ __volatile__ (".byte 0x0f,0xae,0xf8\n\t");
3352#elif RT_INLINE_ASM_USES_INTRIN
3353 _mm_sfence();
3354#else
3355 __asm
3356 {
3357 _emit 0x0f
3358 _emit 0xae
3359 _emit 0xf8
3360 }
3361#endif
3362}
3363
3364
3365/**
3366 * Memory load fence, waits for any pending reads to complete.
3367 * Requires the X86_CPUID_FEATURE_EDX_SSE2 CPUID bit set.
3368 */
3369DECLINLINE(void) ASMReadFenceSSE2(void)
3370{
3371#if RT_INLINE_ASM_GNU_STYLE
3372 __asm__ __volatile__ (".byte 0x0f,0xae,0xe8\n\t");
3373#elif RT_INLINE_ASM_USES_INTRIN
3374 _mm_lfence();
3375#else
3376 __asm
3377 {
3378 _emit 0x0f
3379 _emit 0xae
3380 _emit 0xe8
3381 }
3382#endif
3383}
3384
3385#if !defined(_MSC_VER) || !defined(RT_ARCH_AMD64)
3386
3387/*
3388 * Clear the AC bit in the EFLAGS register.
3389 * Requires the X86_CPUID_STEXT_FEATURE_EBX_SMAP CPUID bit set.
3390 * Requires to be executed in R0.
3391 */
3392DECLINLINE(void) ASMClearAC(void)
3393{
3394#if RT_INLINE_ASM_GNU_STYLE
3395 __asm__ __volatile__ (".byte 0x0f,0x01,0xca\n\t");
3396#else
3397 __asm
3398 {
3399 _emit 0x0f
3400 _emit 0x01
3401 _emit 0xca
3402 }
3403#endif
3404}
3405
3406
3407/*
3408 * Set the AC bit in the EFLAGS register.
3409 * Requires the X86_CPUID_STEXT_FEATURE_EBX_SMAP CPUID bit set.
3410 * Requires to be executed in R0.
3411 */
3412DECLINLINE(void) ASMSetAC(void)
3413{
3414#if RT_INLINE_ASM_GNU_STYLE
3415 __asm__ __volatile__ (".byte 0x0f,0x01,0xcb\n\t");
3416#else
3417 __asm
3418 {
3419 _emit 0x0f
3420 _emit 0x01
3421 _emit 0xcb
3422 }
3423#endif
3424}
3425
3426#endif /* !_MSC_VER || !RT_ARCH_AMD64 */
3427
3428
3429/*
3430 * Include #pragma aux definitions for Watcom C/C++.
3431 */
3432#if defined(__WATCOMC__) && ARCH_BITS == 16
3433# define IPRT_ASM_AMD64_X86_WATCOM_16_INSTANTIATE
3434# undef IPRT_INCLUDED_asm_amd64_x86_watcom_16_h
3435# include "asm-amd64-x86-watcom-16.h"
3436#elif defined(__WATCOMC__) && ARCH_BITS == 32
3437# define IPRT_ASM_AMD64_X86_WATCOM_32_INSTANTIATE
3438# undef IPRT_INCLUDED_asm_amd64_x86_watcom_32_h
3439# include "asm-amd64-x86-watcom-32.h"
3440#endif
3441
3442
3443/** @} */
3444#endif /* !IPRT_INCLUDED_asm_amd64_x86_h */
3445
Note: See TracBrowser for help on using the repository browser.

© 2023 Oracle
Contact – Privacy policy – Terms of Use