asm.h@ 97778

Last change on this file since 97778 was 96407, checked in by vboxsync, 2 years ago
scm copyright and license note update
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 259.8 KB

Line
1	/** @file
2	* IPRT - Assembly Functions.
3	*/
4
5	/*
6	* Copyright (C) 2006-2022 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_h
37	#define IPRT_INCLUDED_asm_h
38	#ifndef RT_WITHOUT_PRAGMA_ONCE
39	# pragma once
40	#endif
41
42	#include <iprt/cdefs.h>
43	#include <iprt/types.h>
44	#include <iprt/assert.h>
45	/** @def RT_INLINE_ASM_USES_INTRIN
46	* Defined as 1 if we're using a _MSC_VER 1400.
47	* Otherwise defined as 0.
48	*/
49
50	/* Solaris 10 header ugliness */
51	#ifdef u
52	# undef u
53	#endif
54
55	#if defined(_MSC_VER) && RT_INLINE_ASM_USES_INTRIN
56	/* Emit the intrinsics at all optimization levels. */
57	# include <iprt/sanitized/intrin.h>
58	# pragma intrinsic(_ReadWriteBarrier)
59	# pragma intrinsic(__cpuid)
60	# pragma intrinsic(__stosd)
61	# pragma intrinsic(__stosw)
62	# pragma intrinsic(__stosb)
63	# pragma intrinsic(_BitScanForward)
64	# pragma intrinsic(_BitScanReverse)
65	# pragma intrinsic(_bittest)
66	# pragma intrinsic(_bittestandset)
67	# pragma intrinsic(_bittestandreset)
68	# pragma intrinsic(_bittestandcomplement)
69	# pragma intrinsic(_byteswap_ushort)
70	# pragma intrinsic(_byteswap_ulong)
71	# pragma intrinsic(_interlockedbittestandset)
72	# pragma intrinsic(_interlockedbittestandreset)
73	# pragma intrinsic(_InterlockedAnd)
74	# pragma intrinsic(_InterlockedOr)
75	# pragma intrinsic(_InterlockedXor)
76	# pragma intrinsic(_InterlockedIncrement)
77	# pragma intrinsic(_InterlockedDecrement)
78	# pragma intrinsic(_InterlockedExchange)
79	# pragma intrinsic(_InterlockedExchangeAdd)
80	# pragma intrinsic(_InterlockedCompareExchange)
81	# pragma intrinsic(_InterlockedCompareExchange8)
82	# pragma intrinsic(_InterlockedCompareExchange16)
83	# pragma intrinsic(_InterlockedCompareExchange64)
84	# pragma intrinsic(_rotl)
85	# pragma intrinsic(_rotr)
86	# pragma intrinsic(_rotl64)
87	# pragma intrinsic(_rotr64)
88	# ifdef RT_ARCH_AMD64
89	# pragma intrinsic(__stosq)
90	# pragma intrinsic(_byteswap_uint64)
91	# pragma intrinsic(_InterlockedCompareExchange128)
92	# pragma intrinsic(_InterlockedExchange64)
93	# pragma intrinsic(_InterlockedExchangeAdd64)
94	# pragma intrinsic(_InterlockedAnd64)
95	# pragma intrinsic(_InterlockedOr64)
96	# pragma intrinsic(_InterlockedIncrement64)
97	# pragma intrinsic(_InterlockedDecrement64)
98	# endif
99	#endif
100
101	/*
102	* Undefine all symbols we have Watcom C/C++ #pragma aux'es for.
103	*/
104	#if defined(__WATCOMC__) && ARCH_BITS == 16 && defined(RT_ARCH_X86)
105	# include "asm-watcom-x86-16.h"
106	#elif defined(__WATCOMC__) && ARCH_BITS == 32 && defined(RT_ARCH_X86)
107	# include "asm-watcom-x86-32.h"
108	#endif
109
110
111	/** @defgroup grp_rt_asm ASM - Assembly Routines
112	* @ingroup grp_rt
113	*
114	* @remarks The difference between ordered and unordered atomic operations are
115	* that the former will complete outstanding reads and writes before
116	* continuing while the latter doesn't make any promises about the
117	* order. Ordered operations doesn't, it seems, make any 100% promise
118	* wrt to whether the operation will complete before any subsequent
119	* memory access. (please, correct if wrong.)
120	*
121	* ASMAtomicSomething operations are all ordered, while
122	* ASMAtomicUoSomething are unordered (note the Uo).
123	*
124	* Please note that ordered operations does not necessarily imply a
125	* compiler (memory) barrier. The user has to use the
126	* ASMCompilerBarrier() macro when that is deemed necessary.
127	*
128	* @remarks Some remarks about __volatile__: Without this keyword gcc is allowed
129	* to reorder or even optimize assembler instructions away. For
130	* instance, in the following code the second rdmsr instruction is
131	* optimized away because gcc treats that instruction as deterministic:
132	*
133	* @code
134	* static inline uint64_t rdmsr_low(int idx)
135	* {
136	* uint32_t low;
137	* __asm__ ("rdmsr" : "=a"(low) : "c"(idx) : "edx");
138	* }
139	* ...
140	* uint32_t msr1 = rdmsr_low(1);
141	* foo(msr1);
142	* msr1 = rdmsr_low(1);
143	* bar(msr1);
144	* @endcode
145	*
146	* The input parameter of rdmsr_low is the same for both calls and
147	* therefore gcc will use the result of the first call as input
148	* parameter for bar() as well. For rdmsr this is not acceptable as
149	* this instruction is _not_ deterministic. This applies to reading
150	* machine status information in general.
151	*
152	* @{
153	*/
154
155
156	/** @def RT_INLINE_ASM_GCC_4_3_X_X86
157	* Used to work around some 4.3.x register allocation issues in this version of
158	* the compiler. So far this workaround is still required for 4.4 and 4.5 but
159	* definitely not for 5.x */
160	#if (RT_GNUC_PREREQ(4, 3) && !RT_GNUC_PREREQ(5, 0) && defined(__i386__))
161	# define RT_INLINE_ASM_GCC_4_3_X_X86 1
162	#else
163	# define RT_INLINE_ASM_GCC_4_3_X_X86 0
164	#endif
165
166	/** @def RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
167	* i686-apple-darwin9-gcc-4.0.1 (GCC) 4.0.1 (Apple Inc. build 5493) screws up
168	* RTSemRWRequestWrite semsemrw-lockless-generic.cpp in release builds. PIC
169	* mode, x86.
170	*
171	* Some gcc 4.3.x versions may have register allocation issues with cmpxchg8b
172	* when in PIC mode on x86.
173	*/
174	#ifndef RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
175	# if defined(DOXYGEN_RUNNING) \|\| defined(__WATCOMC__) /* Watcom has trouble with the expression below */
176	# define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 1
177	# elif defined(_MSC_VER) /* Visual C++ has trouble too, but it'll only tell us when C4688 is enabled. */
178	# define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 0
179	# elif ( (defined(PIC) \|\| defined(__PIC__)) \
180	&& defined(RT_ARCH_X86) \
181	&& ( RT_INLINE_ASM_GCC_4_3_X_X86 \
182	\|\| defined(RT_OS_DARWIN)) )
183	# define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 1
184	# else
185	# define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 0
186	# endif
187	#endif
188
189
190	/** @def RT_INLINE_ASM_EXTERNAL_TMP_ARM
191	* Temporary version of RT_INLINE_ASM_EXTERNAL that excludes ARM. */
192	#if RT_INLINE_ASM_EXTERNAL && !(defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32))
193	# define RT_INLINE_ASM_EXTERNAL_TMP_ARM 1
194	#else
195	# define RT_INLINE_ASM_EXTERNAL_TMP_ARM 0
196	#endif
197
198	/*
199	* ARM is great fun.
200	*/
201	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
202
203	# define RTASM_ARM_NO_BARRIER
204	# ifdef RT_ARCH_ARM64
205	# define RTASM_ARM_NO_BARRIER_IN_REG
206	# define RTASM_ARM_NO_BARRIER_COMMA_IN_REG
207	# define RTASM_ARM_DSB_SY "dsb sy\n\t"
208	# define RTASM_ARM_DSB_SY_IN_REG
209	# define RTASM_ARM_DSB_SY_COMMA_IN_REG
210	# define RTASM_ARM_DMB_SY "dmb sy\n\t"
211	# define RTASM_ARM_DMB_SY_IN_REG
212	# define RTASM_ARM_DMB_SY_COMMA_IN_REG
213	# define RTASM_ARM_DMB_ST "dmb st\n\t"
214	# define RTASM_ARM_DMB_ST_IN_REG
215	# define RTASM_ARM_DMB_ST_COMMA_IN_REG
216	# define RTASM_ARM_DMB_LD "dmb ld\n\t"
217	# define RTASM_ARM_DMB_LD_IN_REG
218	# define RTASM_ARM_DMB_LD_COMMA_IN_REG
219	# define RTASM_ARM_PICK_6432(expr64, expr32) expr64
220	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \
221	uint32_t rcSpill; \
222	uint32_t u32NewRet; \
223	__asm__ __volatile__(".Ltry_again_" #name "_%=:\n\t" \
224	RTASM_ARM_##barrier_type /* before lable? */ \
225	"ldaxr %w[uNew], %[pMem]\n\t" \
226	modify64 \
227	"stlxr %w[rc], %w[uNew], %[pMem]\n\t" \
228	"cbnz %w[rc], .Ltry_again_" #name "_%=\n\t" \
229	: [pMem] "+Q" (*a_pu32Mem) \
230	, [uNew] "=&r" (u32NewRet) \
231	, [rc] "=&r" (rcSpill) \
232	: in_reg \
233	: "cc")
234	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \
235	uint32_t rcSpill; \
236	uint32_t u32OldRet; \
237	uint32_t u32NewSpill; \
238	__asm__ __volatile__(".Ltry_again_" #name "_%=:\n\t" \
239	RTASM_ARM_##barrier_type /* before lable? */ \
240	"ldaxr %w[uOld], %[pMem]\n\t" \
241	modify64 \
242	"stlxr %w[rc], %w[uNew], %[pMem]\n\t" \
243	"cbnz %w[rc], .Ltry_again_" #name "_%=\n\t" \
244	: [pMem] "+Q" (*a_pu32Mem) \
245	, [uOld] "=&r" (u32OldRet) \
246	, [uNew] "=&r" (u32NewSpill) \
247	, [rc] "=&r" (rcSpill) \
248	: in_reg \
249	: "cc")
250	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \
251	uint32_t rcSpill; \
252	uint64_t u64NewRet; \
253	__asm__ __volatile__(".Ltry_again_" #name "_%=:\n\t" \
254	RTASM_ARM_##barrier_type /* before lable? */ \
255	"ldaxr %[uNew], %[pMem]\n\t" \
256	modify64 \
257	"stlxr %w[rc], %[uNew], %[pMem]\n\t" \
258	"cbnz %w[rc], .Ltry_again_" #name "_%=\n\t" \
259	: [pMem] "+Q" (*a_pu64Mem) \
260	, [uNew] "=&r" (u64NewRet) \
261	, [rc] "=&r" (rcSpill) \
262	: in_reg \
263	: "cc")
264	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \
265	uint32_t rcSpill; \
266	uint64_t u64OldRet; \
267	uint64_t u64NewSpill; \
268	__asm__ __volatile__(".Ltry_again_" #name "_%=:\n\t" \
269	RTASM_ARM_##barrier_type /* before lable? */ \
270	"ldaxr %[uOld], %[pMem]\n\t" \
271	modify64 \
272	"stlxr %w[rc], %[uNew], %[pMem]\n\t" \
273	"cbnz %w[rc], .Ltry_again_" #name "_%=\n\t" \
274	: [pMem] "+Q" (*a_pu64Mem) \
275	, [uOld] "=&r" (u64OldRet) \
276	, [uNew] "=&r" (u64NewSpill) \
277	, [rc] "=&r" (rcSpill) \
278	: in_reg \
279	: "cc")
280
281	# else /* RT_ARCH_ARM32 */
282	# define RTASM_ARM_PICK_6432(expr64, expr32) expr32
283	# if RT_ARCH_ARM32 >= 7
284	# warning armv7
285	# define RTASM_ARM_NO_BARRIER_IN_REG
286	# define RTASM_ARM_NO_BARRIER_COMMA_IN_REG
287	# define RTASM_ARM_DSB_SY "dsb sy\n\t"
288	# define RTASM_ARM_DSB_SY_IN_REG "X" (0xfade)
289	# define RTASM_ARM_DMB_SY "dmb sy\n\t"
290	# define RTASM_ARM_DMB_SY_IN_REG "X" (0xfade)
291	# define RTASM_ARM_DMB_ST "dmb st\n\t"
292	# define RTASM_ARM_DMB_ST_IN_REG "X" (0xfade)
293	# define RTASM_ARM_DMB_LD "dmb ld\n\t"
294	# define RTASM_ARM_DMB_LD_IN_REG "X" (0xfade)
295
296	# elif RT_ARCH_ARM32 >= 6
297	# warning armv6
298	# define RTASM_ARM_DSB_SY "mcr p15, 0, %[uZero], c7, c10, 4\n\t"
299	# define RTASM_ARM_DSB_SY_IN_REG [uZero] "r" (0)
300	# define RTASM_ARM_DMB_SY "mcr p15, 0, %[uZero], c7, c10, 5\n\t"
301	# define RTASM_ARM_DMB_SY_IN_REG [uZero] "r" (0)
302	# define RTASM_ARM_DMB_ST RTASM_ARM_DMB_SY
303	# define RTASM_ARM_DMB_ST_IN_REG RTASM_ARM_DMB_SY_IN_REG
304	# define RTASM_ARM_DMB_LD RTASM_ARM_DMB_SY
305	# define RTASM_ARM_DMB_LD_IN_REG RTASM_ARM_DMB_SY_IN_REG
306	# elif RT_ARCH_ARM32 >= 4
307	# warning armv5 or older
308	# define RTASM_ARM_DSB_SY "mcr p15, 0, %[uZero], c7, c10, 4\n\t"
309	# define RTASM_ARM_DSB_SY_IN_REG [uZero] "r" (0)
310	# define RTASM_ARM_DMB_SY RTASM_ARM_DSB_SY
311	# define RTASM_ARM_DMB_SY_IN_REG RTASM_ARM_DSB_SY_IN_REG
312	# define RTASM_ARM_DMB_ST RTASM_ARM_DSB_SY
313	# define RTASM_ARM_DMB_ST_IN_REG RTASM_ARM_DSB_SY_IN_REG
314	# define RTASM_ARM_DMB_LD RTASM_ARM_DSB_SY
315	# define RTASM_ARM_DMB_LD_IN_REG RTASM_ARM_DSB_SY_IN_REG
316	# else
317	# error "huh? Odd RT_ARCH_ARM32 value!"
318	# endif
319	# define RTASM_ARM_DSB_SY_COMMA_IN_REG , RTASM_ARM_DSB_SY_IN_REG
320	# define RTASM_ARM_DMB_SY_COMMA_IN_REG , RTASM_ARM_DMB_SY_IN_REG
321	# define RTASM_ARM_DMB_ST_COMMA_IN_REG , RTASM_ARM_DMB_ST_IN_REG
322	# define RTASM_ARM_DMB_LD_COMMA_IN_REG , RTASM_ARM_DMB_LD_IN_REG
323	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \
324	uint32_t rcSpill; \
325	uint32_t u32NewRet; \
326	__asm__ __volatile__(".Ltry_again_" #name "_%=:\n\t" \
327	RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \
328	"ldrex %[uNew], %[pMem]\n\t" \
329	modify32 \
330	"strex %[rc], %[uNew], %[pMem]\n\t" \
331	"cmp %[rc], #0\n\t" \
332	"bne .Ltry_again_" #name "_%=\n\t" \
333	: [pMem] "+m" (*a_pu32Mem) \
334	, [uNew] "=&r" (u32NewRet) \
335	, [rc] "=&r" (rcSpill) \
336	: RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \
337	, in_reg \
338	: "cc")
339	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \
340	uint32_t rcSpill; \
341	uint32_t u32OldRet; \
342	uint32_t u32NewSpill; \
343	__asm__ __volatile__(".Ltry_again_" #name "_%=:\n\t" \
344	RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \
345	"ldrex %[uOld], %[pMem]\n\t" \
346	modify32 \
347	"strex %[rc], %[uNew], %[pMem]\n\t" \
348	"cmp %[rc], #0\n\t" \
349	"bne .Ltry_again_" #name "_%=\n\t" \
350	: [pMem] "+m" (*a_pu32Mem) \
351	, [uOld] "=&r" (u32OldRet) \
352	, [uNew] "=&r" (u32NewSpill) \
353	, [rc] "=&r" (rcSpill) \
354	: RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \
355	, in_reg \
356	: "cc")
357	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \
358	uint32_t rcSpill; \
359	uint64_t u64NewRet; \
360	__asm__ __volatile__(".Ltry_again_" #name "_%=:\n\t" \
361	RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \
362	"ldrexd %[uNew], %H[uNew], %[pMem]\n\t" \
363	modify32 \
364	"strexd %[rc], %[uNew], %H[uNew], %[pMem]\n\t" \
365	"cmp %[rc], #0\n\t" \
366	"bne .Ltry_again_" #name "_%=\n\t" \
367	: [pMem] "+m" (*a_pu64Mem), \
368	[uNew] "=&r" (u64NewRet), \
369	[rc] "=&r" (rcSpill) \
370	: RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \
371	, in_reg \
372	: "cc")
373	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \
374	uint32_t rcSpill; \
375	uint64_t u64OldRet; \
376	uint64_t u64NewSpill; \
377	__asm__ __volatile__(".Ltry_again_" #name "_%=:\n\t" \
378	RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \
379	"ldrexd %[uOld], %H[uOld], %[pMem]\n\t" \
380	modify32 \
381	"strexd %[rc], %[uNew], %H[uNew], %[pMem]\n\t" \
382	"cmp %[rc], #0\n\t" \
383	"bne .Ltry_again_" #name "_%=\n\t" \
384	: [pMem] "+m" (*a_pu64Mem), \
385	[uOld] "=&r" (u64OldRet), \
386	[uNew] "=&r" (u64NewSpill), \
387	[rc] "=&r" (rcSpill) \
388	: RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \
389	, in_reg \
390	: "cc")
391	# endif /* RT_ARCH_ARM32 */
392	#endif
393
394
395	/** @def ASMReturnAddress
396	* Gets the return address of the current (or calling if you like) function or method.
397	*/
398	#ifdef _MSC_VER
399	# ifdef __cplusplus
400	extern "C"
401	# endif
402	void * _ReturnAddress(void);
403	# pragma intrinsic(_ReturnAddress)
404	# define ASMReturnAddress() _ReturnAddress()
405	#elif defined(__GNUC__) \|\| defined(DOXYGEN_RUNNING)
406	# define ASMReturnAddress() __builtin_return_address(0)
407	#elif defined(__WATCOMC__)
408	# define ASMReturnAddress() Watcom_does_not_appear_to_have_intrinsic_return_address_function()
409	#else
410	# error "Unsupported compiler."
411	#endif
412
413
414	/**
415	* Compiler memory barrier.
416	*
417	* Ensure that the compiler does not use any cached (register/tmp stack) memory
418	* values or any outstanding writes when returning from this function.
419	*
420	* This function must be used if non-volatile data is modified by a
421	* device or the VMM. Typical cases are port access, MMIO access,
422	* trapping instruction, etc.
423	*/
424	#if RT_INLINE_ASM_GNU_STYLE
425	# define ASMCompilerBarrier() do { __asm__ __volatile__("" : : : "memory"); } while (0)
426	#elif RT_INLINE_ASM_USES_INTRIN
427	# define ASMCompilerBarrier() do { _ReadWriteBarrier(); } while (0)
428	#elif defined(__WATCOMC__)
429	void ASMCompilerBarrier(void);
430	#else /* 2003 should have _ReadWriteBarrier() but I guess we're at 2002 level then... */
431	DECLINLINE(void) ASMCompilerBarrier(void) RT_NOTHROW_DEF
432	{
433	__asm
434	{
435	}
436	}
437	#endif
438
439
440	/** @def ASMBreakpoint
441	* Debugger Breakpoint.
442	* @deprecated Use RT_BREAKPOINT instead.
443	* @internal
444	*/
445	#define ASMBreakpoint() RT_BREAKPOINT()
446
447
448	/**
449	* Spinloop hint for platforms that have these, empty function on the other
450	* platforms.
451	*
452	* x86 & AMD64: The PAUSE variant of NOP for helping hyperthreaded CPUs detecting
453	* spin locks.
454	*/
455	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
456	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMNopPause(void) RT_NOTHROW_PROTO;
457	#else
458	DECLINLINE(void) ASMNopPause(void) RT_NOTHROW_DEF
459	{
460	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
461	# if RT_INLINE_ASM_GNU_STYLE
462	__asm__ __volatile__(".byte 0xf3,0x90\n\t");
463	# else
464	__asm {
465	_emit 0f3h
466	_emit 090h
467	}
468	# endif
469
470	# elif defined(RT_ARCH_ARM32) \|\| defined(RT_ARCH_ARM64)
471	__asm__ __volatile__("yield\n\t"); /* ARMv6K+ */
472
473	# else
474	/* dummy */
475	# endif
476	}
477	#endif
478
479
480	/**
481	* Atomically Exchange an unsigned 8-bit value, ordered.
482	*
483	* @returns Current *pu8 value
484	* @param pu8 Pointer to the 8-bit variable to update.
485	* @param u8 The 8-bit value to assign to *pu8.
486	*/
487	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
488	RT_ASM_DECL_PRAGMA_WATCOM(uint8_t) ASMAtomicXchgU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_PROTO;
489	#else
490	DECLINLINE(uint8_t) ASMAtomicXchgU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_DEF
491	{
492	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
493	# if RT_INLINE_ASM_GNU_STYLE
494	__asm__ __volatile__("xchgb %0, %1\n\t"
495	: "=m" (*pu8)
496	, "=q" (u8) /* =r - busted on g++ (GCC) 3.4.4 20050721 (Red Hat 3.4.4-2) */
497	: "1" (u8)
498	, "m" (*pu8));
499	# else
500	__asm
501	{
502	# ifdef RT_ARCH_AMD64
503	mov rdx, [pu8]
504	mov al, [u8]
505	xchg [rdx], al
506	mov [u8], al
507	# else
508	mov edx, [pu8]
509	mov al, [u8]
510	xchg [edx], al
511	mov [u8], al
512	# endif
513	}
514	# endif
515	return u8;
516
517	# elif defined(RT_ARCH_ARM32) \|\| defined(RT_ARCH_ARM64)
518	uint32_t uOld;
519	uint32_t rcSpill;
520	__asm__ __volatile__(".Ltry_again_ASMAtomicXchgU8_%=:\n\t"
521	RTASM_ARM_DMB_SY
522	# if defined(RT_ARCH_ARM64)
523	"ldaxrb %w[uOld], %[pMem]\n\t"
524	"stlxrb %w[rc], %w[uNew], %[pMem]\n\t"
525	"cbnz %w[rc], .Ltry_again_ASMAtomicXchgU8_%=\n\t"
526	# else
527	"ldrexb %[uOld], %[pMem]\n\t" /* ARMv6+ */
528	"strexb %[rc], %[uNew], %[pMem]\n\t"
529	"cmp %[rc], #0\n\t"
530	"bne .Ltry_again_ASMAtomicXchgU8_%=\n\t"
531	# endif
532	: [pMem] "+Q" (*pu8)
533	, [uOld] "=&r" (uOld)
534	, [rc] "=&r" (rcSpill)
535	: [uNew] "r" ((uint32_t)u8)
536	RTASM_ARM_DMB_SY_COMMA_IN_REG
537	: "cc");
538	return (uint8_t)uOld;
539
540	# else
541	# error "Port me"
542	# endif
543	}
544	#endif
545
546
547	/**
548	* Atomically Exchange a signed 8-bit value, ordered.
549	*
550	* @returns Current *pu8 value
551	* @param pi8 Pointer to the 8-bit variable to update.
552	* @param i8 The 8-bit value to assign to *pi8.
553	*/
554	DECLINLINE(int8_t) ASMAtomicXchgS8(volatile int8_t RT_FAR *pi8, int8_t i8) RT_NOTHROW_DEF
555	{
556	return (int8_t)ASMAtomicXchgU8((volatile uint8_t RT_FAR *)pi8, (uint8_t)i8);
557	}
558
559
560	/**
561	* Atomically Exchange a bool value, ordered.
562	*
563	* @returns Current *pf value
564	* @param pf Pointer to the 8-bit variable to update.
565	* @param f The 8-bit value to assign to *pi8.
566	*/
567	DECLINLINE(bool) ASMAtomicXchgBool(volatile bool RT_FAR *pf, bool f) RT_NOTHROW_DEF
568	{
569	#ifdef _MSC_VER
570	return !!ASMAtomicXchgU8((volatile uint8_t RT_FAR *)pf, (uint8_t)f);
571	#else
572	return (bool)ASMAtomicXchgU8((volatile uint8_t RT_FAR *)pf, (uint8_t)f);
573	#endif
574	}
575
576
577	/**
578	* Atomically Exchange an unsigned 16-bit value, ordered.
579	*
580	* @returns Current *pu16 value
581	* @param pu16 Pointer to the 16-bit variable to update.
582	* @param u16 The 16-bit value to assign to *pu16.
583	*/
584	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
585	RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMAtomicXchgU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_PROTO;
586	#else
587	DECLINLINE(uint16_t) ASMAtomicXchgU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_DEF
588	{
589	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
590	# if RT_INLINE_ASM_GNU_STYLE
591	__asm__ __volatile__("xchgw %0, %1\n\t"
592	: "=m" (*pu16)
593	, "=r" (u16)
594	: "1" (u16)
595	, "m" (*pu16));
596	# else
597	__asm
598	{
599	# ifdef RT_ARCH_AMD64
600	mov rdx, [pu16]
601	mov ax, [u16]
602	xchg [rdx], ax
603	mov [u16], ax
604	# else
605	mov edx, [pu16]
606	mov ax, [u16]
607	xchg [edx], ax
608	mov [u16], ax
609	# endif
610	}
611	# endif
612	return u16;
613
614	# elif defined(RT_ARCH_ARM32) \|\| defined(RT_ARCH_ARM64)
615	uint32_t uOld;
616	uint32_t rcSpill;
617	__asm__ __volatile__(".Ltry_again_ASMAtomicXchgU16_%=:\n\t"
618	RTASM_ARM_DMB_SY
619	# if defined(RT_ARCH_ARM64)
620	"ldaxrh %w[uOld], %[pMem]\n\t"
621	"stlxrh %w[rc], %w[uNew], %[pMem]\n\t"
622	"cbnz %w[rc], .Ltry_again_ASMAtomicXchgU16_%=\n\t"
623	# else
624	"ldrexh %[uOld], %[pMem]\n\t" /* ARMv6+ */
625	"strexh %[rc], %[uNew], %[pMem]\n\t"
626	"cmp %[rc], #0\n\t"
627	"bne .Ltry_again_ASMAtomicXchgU16_%=\n\t"
628	# endif
629	: [pMem] "+Q" (*pu16)
630	, [uOld] "=&r" (uOld)
631	, [rc] "=&r" (rcSpill)
632	: [uNew] "r" ((uint32_t)u16)
633	RTASM_ARM_DMB_SY_COMMA_IN_REG
634	: "cc");
635	return (uint16_t)uOld;
636
637	# else
638	# error "Port me"
639	# endif
640	}
641	#endif
642
643
644	/**
645	* Atomically Exchange a signed 16-bit value, ordered.
646	*
647	* @returns Current *pu16 value
648	* @param pi16 Pointer to the 16-bit variable to update.
649	* @param i16 The 16-bit value to assign to *pi16.
650	*/
651	DECLINLINE(int16_t) ASMAtomicXchgS16(volatile int16_t RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF
652	{
653	return (int16_t)ASMAtomicXchgU16((volatile uint16_t RT_FAR *)pi16, (uint16_t)i16);
654	}
655
656
657	/**
658	* Atomically Exchange an unsigned 32-bit value, ordered.
659	*
660	* @returns Current *pu32 value
661	* @param pu32 Pointer to the 32-bit variable to update.
662	* @param u32 The 32-bit value to assign to *pu32.
663	*
664	* @remarks Does not work on 286 and earlier.
665	*/
666	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
667	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicXchgU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
668	#else
669	DECLINLINE(uint32_t) ASMAtomicXchgU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
670	{
671	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
672	# if RT_INLINE_ASM_GNU_STYLE
673	__asm__ __volatile__("xchgl %0, %1\n\t"
674	: "=m" (pu32) /* @todo r=bird: +m rather than =m here? */
675	, "=r" (u32)
676	: "1" (u32)
677	, "m" (*pu32));
678
679	# elif RT_INLINE_ASM_USES_INTRIN
680	u32 = _InterlockedExchange((long RT_FAR *)pu32, u32);
681
682	# else
683	__asm
684	{
685	# ifdef RT_ARCH_AMD64
686	mov rdx, [pu32]
687	mov eax, u32
688	xchg [rdx], eax
689	mov [u32], eax
690	# else
691	mov edx, [pu32]
692	mov eax, u32
693	xchg [edx], eax
694	mov [u32], eax
695	# endif
696	}
697	# endif
698	return u32;
699
700	# elif defined(RT_ARCH_ARM32) \|\| defined(RT_ARCH_ARM64)
701	uint32_t uOld;
702	uint32_t rcSpill;
703	__asm__ __volatile__(".Ltry_again_ASMAtomicXchgU32_%=:\n\t"
704	RTASM_ARM_DMB_SY
705	# if defined(RT_ARCH_ARM64)
706	"ldaxr %w[uOld], %[pMem]\n\t"
707	"stlxr %w[rc], %w[uNew], %[pMem]\n\t"
708	"cbnz %w[rc], .Ltry_again_ASMAtomicXchgU32_%=\n\t"
709	# else
710	"ldrex %[uOld], %[pMem]\n\t" /* ARMv6+ */
711	"strex %[rc], %[uNew], %[pMem]\n\t"
712	"cmp %[rc], #0\n\t"
713	"bne .Ltry_again_ASMAtomicXchgU32_%=\n\t"
714	# endif
715	: [pMem] "+Q" (*pu32)
716	, [uOld] "=&r" (uOld)
717	, [rc] "=&r" (rcSpill)
718	: [uNew] "r" (u32)
719	RTASM_ARM_DMB_SY_COMMA_IN_REG
720	: "cc");
721	return uOld;
722
723	# else
724	# error "Port me"
725	# endif
726	}
727	#endif
728
729
730	/**
731	* Atomically Exchange a signed 32-bit value, ordered.
732	*
733	* @returns Current *pu32 value
734	* @param pi32 Pointer to the 32-bit variable to update.
735	* @param i32 The 32-bit value to assign to *pi32.
736	*/
737	DECLINLINE(int32_t) ASMAtomicXchgS32(volatile int32_t RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
738	{
739	return (int32_t)ASMAtomicXchgU32((volatile uint32_t RT_FAR *)pi32, (uint32_t)i32);
740	}
741
742
743	/**
744	* Atomically Exchange an unsigned 64-bit value, ordered.
745	*
746	* @returns Current *pu64 value
747	* @param pu64 Pointer to the 64-bit variable to update.
748	* @param u64 The 64-bit value to assign to *pu64.
749	*
750	* @remarks Works on 32-bit x86 CPUs starting with Pentium.
751	*/
752	#if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \
753	\|\| RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
754	RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicXchgU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
755	#else
756	DECLINLINE(uint64_t) ASMAtomicXchgU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
757	{
758	# if defined(RT_ARCH_AMD64)
759	# if RT_INLINE_ASM_USES_INTRIN
760	return _InterlockedExchange64((__int64 *)pu64, u64);
761
762	# elif RT_INLINE_ASM_GNU_STYLE
763	__asm__ __volatile__("xchgq %0, %1\n\t"
764	: "=m" (*pu64)
765	, "=r" (u64)
766	: "1" (u64)
767	, "m" (*pu64));
768	return u64;
769	# else
770	__asm
771	{
772	mov rdx, [pu64]
773	mov rax, [u64]
774	xchg [rdx], rax
775	mov [u64], rax
776	}
777	return u64;
778	# endif
779
780	# elif defined(RT_ARCH_X86)
781	# if RT_INLINE_ASM_GNU_STYLE
782	# if defined(PIC) \|\| defined(__PIC__)
783	uint32_t u32EBX = (uint32_t)u64;
784	__asm__ __volatile__(/"xchgl %%esi, %5\n\t"/
785	"xchgl %%ebx, %3\n\t"
786	"1:\n\t"
787	"lock; cmpxchg8b (%5)\n\t"
788	"jnz 1b\n\t"
789	"movl %3, %%ebx\n\t"
790	/"xchgl %%esi, %5\n\t"/
791	: "=A" (u64)
792	, "=m" (*pu64)
793	: "0" (*pu64)
794	, "m" ( u32EBX )
795	, "c" ( (uint32_t)(u64 >> 32) )
796	, "S" (pu64)
797	: "cc");
798	# else /* !PIC */
799	__asm__ __volatile__("1:\n\t"
800	"lock; cmpxchg8b %1\n\t"
801	"jnz 1b\n\t"
802	: "=A" (u64)
803	, "=m" (*pu64)
804	: "0" (*pu64)
805	, "b" ( (uint32_t)u64 )
806	, "c" ( (uint32_t)(u64 >> 32) )
807	: "cc");
808	# endif
809	# else
810	__asm
811	{
812	mov ebx, dword ptr [u64]
813	mov ecx, dword ptr [u64 + 4]
814	mov edi, pu64
815	mov eax, dword ptr [edi]
816	mov edx, dword ptr [edi + 4]
817	retry:
818	lock cmpxchg8b [edi]
819	jnz retry
820	mov dword ptr [u64], eax
821	mov dword ptr [u64 + 4], edx
822	}
823	# endif
824	return u64;
825
826	# elif defined(RT_ARCH_ARM32) \|\| defined(RT_ARCH_ARM64)
827	uint32_t rcSpill;
828	uint64_t uOld;
829	__asm__ __volatile__(".Ltry_again_ASMAtomicXchgU64_%=:\n\t"
830	RTASM_ARM_DMB_SY
831	# if defined(RT_ARCH_ARM64)
832	"ldaxr %[uOld], %[pMem]\n\t"
833	"stlxr %w[rc], %[uNew], %[pMem]\n\t"
834	"cbnz %w[rc], .Ltry_again_ASMAtomicXchgU64_%=\n\t"
835	# else
836	"ldrexd %[uOld], %H[uOld], %[pMem]\n\t" /* ARMv6+ */
837	"strexd %[rc], %[uNew], %H[uNew], %[pMem]\n\t"
838	"cmp %[rc], #0\n\t"
839	"bne .Ltry_again_ASMAtomicXchgU64_%=\n\t"
840	# endif
841	: [pMem] "+Q" (*pu64)
842	, [uOld] "=&r" (uOld)
843	, [rc] "=&r" (rcSpill)
844	: [uNew] "r" (u64)
845	RTASM_ARM_DMB_SY_COMMA_IN_REG
846	: "cc");
847	return uOld;
848
849	# else
850	# error "Port me"
851	# endif
852	}
853	#endif
854
855
856	/**
857	* Atomically Exchange an signed 64-bit value, ordered.
858	*
859	* @returns Current *pi64 value
860	* @param pi64 Pointer to the 64-bit variable to update.
861	* @param i64 The 64-bit value to assign to *pi64.
862	*/
863	DECLINLINE(int64_t) ASMAtomicXchgS64(volatile int64_t RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
864	{
865	return (int64_t)ASMAtomicXchgU64((volatile uint64_t RT_FAR *)pi64, (uint64_t)i64);
866	}
867
868
869	/**
870	* Atomically Exchange a size_t value, ordered.
871	*
872	* @returns Current *ppv value
873	* @param puDst Pointer to the size_t variable to update.
874	* @param uNew The new value to assign to *puDst.
875	*/
876	DECLINLINE(size_t) ASMAtomicXchgZ(size_t volatile RT_FAR *puDst, const size_t uNew) RT_NOTHROW_DEF
877	{
878	#if ARCH_BITS == 16
879	AssertCompile(sizeof(size_t) == 2);
880	return ASMAtomicXchgU16((volatile uint16_t RT_FAR *)puDst, uNew);
881	#elif ARCH_BITS == 32
882	return ASMAtomicXchgU32((volatile uint32_t RT_FAR *)puDst, uNew);
883	#elif ARCH_BITS == 64
884	return ASMAtomicXchgU64((volatile uint64_t RT_FAR *)puDst, uNew);
885	#else
886	# error "ARCH_BITS is bogus"
887	#endif
888	}
889
890
891	/**
892	* Atomically Exchange a pointer value, ordered.
893	*
894	* @returns Current *ppv value
895	* @param ppv Pointer to the pointer variable to update.
896	* @param pv The pointer value to assign to *ppv.
897	*/
898	DECLINLINE(void RT_FAR ) ASMAtomicXchgPtr(void RT_FAR volatile RT_FAR ppv, const void RT_FAR pv) RT_NOTHROW_DEF
899	{
900	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
901	return (void RT_FAR )ASMAtomicXchgU32((volatile uint32_t RT_FAR )(void RT_FAR *)ppv, (uint32_t)pv);
902	#elif ARCH_BITS == 64
903	return (void RT_FAR )ASMAtomicXchgU64((volatile uint64_t RT_FAR )(void RT_FAR *)ppv, (uint64_t)pv);
904	#else
905	# error "ARCH_BITS is bogus"
906	#endif
907	}
908
909
910	/**
911	* Convenience macro for avoiding the annoying casting with ASMAtomicXchgPtr.
912	*
913	* @returns Current *pv value
914	* @param ppv Pointer to the pointer variable to update.
915	* @param pv The pointer value to assign to *ppv.
916	* @param Type The type of *ppv, sans volatile.
917	*/
918	#ifdef __GNUC__ /* 8.2.0 requires -Wno-ignored-qualifiers */
919	# define ASMAtomicXchgPtrT(ppv, pv, Type) \
920	__extension__ \
921	({\
922	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
923	Type const pvTypeChecked = (pv); \
924	Type pvTypeCheckedRet = (__typeof__((ppv))) ASMAtomicXchgPtr((void volatile )ppvTypeChecked, (void )pvTypeChecked); \
925	pvTypeCheckedRet; \
926	})
927	#else
928	# define ASMAtomicXchgPtrT(ppv, pv, Type) \
929	(Type)ASMAtomicXchgPtr((void RT_FAR * volatile RT_FAR )(ppv), (void RT_FAR )(pv))
930	#endif
931
932
933	/**
934	* Atomically Exchange a raw-mode context pointer value, ordered.
935	*
936	* @returns Current *ppv value
937	* @param ppvRC Pointer to the pointer variable to update.
938	* @param pvRC The pointer value to assign to *ppv.
939	*/
940	DECLINLINE(RTRCPTR) ASMAtomicXchgRCPtr(RTRCPTR volatile RT_FAR *ppvRC, RTRCPTR pvRC) RT_NOTHROW_DEF
941	{
942	return (RTRCPTR)ASMAtomicXchgU32((uint32_t volatile RT_FAR )(void RT_FAR )ppvRC, (uint32_t)pvRC);
943	}
944
945
946	/**
947	* Atomically Exchange a ring-0 pointer value, ordered.
948	*
949	* @returns Current *ppv value
950	* @param ppvR0 Pointer to the pointer variable to update.
951	* @param pvR0 The pointer value to assign to *ppv.
952	*/
953	DECLINLINE(RTR0PTR) ASMAtomicXchgR0Ptr(RTR0PTR volatile RT_FAR *ppvR0, RTR0PTR pvR0) RT_NOTHROW_DEF
954	{
955	#if R0_ARCH_BITS == 32 \|\| ARCH_BITS == 16
956	return (RTR0PTR)ASMAtomicXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )ppvR0, (uint32_t)pvR0);
957	#elif R0_ARCH_BITS == 64
958	return (RTR0PTR)ASMAtomicXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )ppvR0, (uint64_t)pvR0);
959	#else
960	# error "R0_ARCH_BITS is bogus"
961	#endif
962	}
963
964
965	/**
966	* Atomically Exchange a ring-3 pointer value, ordered.
967	*
968	* @returns Current *ppv value
969	* @param ppvR3 Pointer to the pointer variable to update.
970	* @param pvR3 The pointer value to assign to *ppv.
971	*/
972	DECLINLINE(RTR3PTR) ASMAtomicXchgR3Ptr(RTR3PTR volatile RT_FAR *ppvR3, RTR3PTR pvR3) RT_NOTHROW_DEF
973	{
974	#if R3_ARCH_BITS == 32 \|\| ARCH_BITS == 16
975	return (RTR3PTR)ASMAtomicXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )ppvR3, (uint32_t)pvR3);
976	#elif R3_ARCH_BITS == 64
977	return (RTR3PTR)ASMAtomicXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )ppvR3, (uint64_t)pvR3);
978	#else
979	# error "R3_ARCH_BITS is bogus"
980	#endif
981	}
982
983
984	/** @def ASMAtomicXchgHandle
985	* Atomically Exchange a typical IPRT handle value, ordered.
986	*
987	* @param ph Pointer to the value to update.
988	* @param hNew The new value to assigned to *pu.
989	* @param phRes Where to store the current *ph value.
990	*
991	* @remarks This doesn't currently work for all handles (like RTFILE).
992	*/
993	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
994	# define ASMAtomicXchgHandle(ph, hNew, phRes) \
995	do { \
996	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
997	AssertCompile(sizeof(*(phRes)) == sizeof(uint32_t)); \
998	(uint32_t RT_FAR )(phRes) = ASMAtomicXchgU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)(hNew)); \
999	} while (0)
1000	#elif HC_ARCH_BITS == 64
1001	# define ASMAtomicXchgHandle(ph, hNew, phRes) \
1002	do { \
1003	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
1004	AssertCompile(sizeof(*(phRes)) == sizeof(uint64_t)); \
1005	(uint64_t RT_FAR )(phRes) = ASMAtomicXchgU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)(hNew)); \
1006	} while (0)
1007	#else
1008	# error HC_ARCH_BITS
1009	#endif
1010
1011
1012	/**
1013	* Atomically Exchange a value which size might differ
1014	* between platforms or compilers, ordered.
1015	*
1016	* @param pu Pointer to the variable to update.
1017	* @param uNew The value to assign to *pu.
1018	* @todo This is busted as its missing the result argument.
1019	*/
1020	#define ASMAtomicXchgSize(pu, uNew) \
1021	do { \
1022	switch (sizeof(*(pu))) { \
1023	case 1: ASMAtomicXchgU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu), (uint8_t)(uNew)); break; \
1024	case 2: ASMAtomicXchgU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu), (uint16_t)(uNew)); break; \
1025	case 4: ASMAtomicXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
1026	case 8: ASMAtomicXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
1027	default: AssertMsgFailed(("ASMAtomicXchgSize: size %d is not supported\n", sizeof(*(pu)))); \
1028	} \
1029	} while (0)
1030
1031	/**
1032	* Atomically Exchange a value which size might differ
1033	* between platforms or compilers, ordered.
1034	*
1035	* @param pu Pointer to the variable to update.
1036	* @param uNew The value to assign to *pu.
1037	* @param puRes Where to store the current *pu value.
1038	*/
1039	#define ASMAtomicXchgSizeCorrect(pu, uNew, puRes) \
1040	do { \
1041	switch (sizeof(*(pu))) { \
1042	case 1: (uint8_t RT_FAR )(puRes) = ASMAtomicXchgU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu), (uint8_t)(uNew)); break; \
1043	case 2: (uint16_t RT_FAR )(puRes) = ASMAtomicXchgU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu), (uint16_t)(uNew)); break; \
1044	case 4: (uint32_t RT_FAR )(puRes) = ASMAtomicXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
1045	case 8: (uint64_t RT_FAR )(puRes) = ASMAtomicXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
1046	default: AssertMsgFailed(("ASMAtomicXchgSize: size %d is not supported\n", sizeof(*(pu)))); \
1047	} \
1048	} while (0)
1049
1050
1051
1052	/**
1053	* Atomically Compare and Exchange an unsigned 8-bit value, ordered.
1054	*
1055	* @returns true if xchg was done.
1056	* @returns false if xchg wasn't done.
1057	*
1058	* @param pu8 Pointer to the value to update.
1059	* @param u8New The new value to assigned to *pu8.
1060	* @param u8Old The old value to *pu8 compare with.
1061	*
1062	* @remarks x86: Requires a 486 or later.
1063	* @todo Rename ASMAtomicCmpWriteU8
1064	*/
1065	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM \|\| !RT_INLINE_ASM_GNU_STYLE
1066	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgU8(volatile uint8_t RT_FAR *pu8, const uint8_t u8New, const uint8_t u8Old) RT_NOTHROW_PROTO;
1067	#else
1068	DECLINLINE(bool) ASMAtomicCmpXchgU8(volatile uint8_t RT_FAR *pu8, const uint8_t u8New, uint8_t u8Old) RT_NOTHROW_DEF
1069	{
1070	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1071	uint8_t u8Ret;
1072	__asm__ __volatile__("lock; cmpxchgb %3, %0\n\t"
1073	"setz %1\n\t"
1074	: "=m" (*pu8)
1075	, "=qm" (u8Ret)
1076	, "=a" (u8Old)
1077	: "q" (u8New)
1078	, "2" (u8Old)
1079	, "m" (*pu8)
1080	: "cc");
1081	return (bool)u8Ret;
1082
1083	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
1084	union { uint32_t u; bool f; } fXchg;
1085	uint32_t u32Spill;
1086	uint32_t rcSpill;
1087	__asm__ __volatile__(".Ltry_again_ASMAtomicCmpXchgU8_%=:\n\t"
1088	RTASM_ARM_DMB_SY
1089	# if defined(RT_ARCH_ARM64)
1090	"ldaxrb %w[uOld], %[pMem]\n\t"
1091	"cmp %w[uOld], %w[uCmp]\n\t"
1092	"bne 1f\n\t" /* stop here if not equal */
1093	"stlxrb %w[rc], %w[uNew], %[pMem]\n\t"
1094	"cbnz %w[rc], .Ltry_again_ASMAtomicCmpXchgU8_%=\n\t"
1095	"mov %w[fXchg], #1\n\t"
1096	# else
1097	"ldrexb %[uOld], %[pMem]\n\t"
1098	"teq %[uOld], %[uCmp]\n\t"
1099	"strexbeq %[rc], %[uNew], %[pMem]\n\t"
1100	"bne 1f\n\t" /* stop here if not equal */
1101	"cmp %[rc], #0\n\t"
1102	"bne .Ltry_again_ASMAtomicCmpXchgU8_%=\n\t"
1103	"mov %[fXchg], #1\n\t"
1104	# endif
1105	"1:\n\t"
1106	: [pMem] "+Q" (*pu8)
1107	, [uOld] "=&r" (u32Spill)
1108	, [rc] "=&r" (rcSpill)
1109	, [fXchg] "=&r" (fXchg.u)
1110	: [uCmp] "r" ((uint32_t)u8Old)
1111	, [uNew] "r" ((uint32_t)u8New)
1112	, "[fXchg]" (0)
1113	RTASM_ARM_DMB_SY_COMMA_IN_REG
1114	: "cc");
1115	return fXchg.f;
1116
1117	# else
1118	# error "Port me"
1119	# endif
1120	}
1121	#endif
1122
1123
1124	/**
1125	* Atomically Compare and Exchange a signed 8-bit value, ordered.
1126	*
1127	* @returns true if xchg was done.
1128	* @returns false if xchg wasn't done.
1129	*
1130	* @param pi8 Pointer to the value to update.
1131	* @param i8New The new value to assigned to *pi8.
1132	* @param i8Old The old value to *pi8 compare with.
1133	*
1134	* @remarks x86: Requires a 486 or later.
1135	* @todo Rename ASMAtomicCmpWriteS8
1136	*/
1137	DECLINLINE(bool) ASMAtomicCmpXchgS8(volatile int8_t RT_FAR *pi8, const int8_t i8New, const int8_t i8Old) RT_NOTHROW_DEF
1138	{
1139	return ASMAtomicCmpXchgU8((volatile uint8_t RT_FAR *)pi8, (uint8_t)i8New, (uint8_t)i8Old);
1140	}
1141
1142
1143	/**
1144	* Atomically Compare and Exchange a bool value, ordered.
1145	*
1146	* @returns true if xchg was done.
1147	* @returns false if xchg wasn't done.
1148	*
1149	* @param pf Pointer to the value to update.
1150	* @param fNew The new value to assigned to *pf.
1151	* @param fOld The old value to *pf compare with.
1152	*
1153	* @remarks x86: Requires a 486 or later.
1154	* @todo Rename ASMAtomicCmpWriteBool
1155	*/
1156	DECLINLINE(bool) ASMAtomicCmpXchgBool(volatile bool RT_FAR *pf, const bool fNew, const bool fOld) RT_NOTHROW_DEF
1157	{
1158	return ASMAtomicCmpXchgU8((volatile uint8_t RT_FAR *)pf, (uint8_t)fNew, (uint8_t)fOld);
1159	}
1160
1161
1162	/**
1163	* Atomically Compare and Exchange an unsigned 32-bit value, ordered.
1164	*
1165	* @returns true if xchg was done.
1166	* @returns false if xchg wasn't done.
1167	*
1168	* @param pu32 Pointer to the value to update.
1169	* @param u32New The new value to assigned to *pu32.
1170	* @param u32Old The old value to *pu32 compare with.
1171	*
1172	* @remarks x86: Requires a 486 or later.
1173	* @todo Rename ASMAtomicCmpWriteU32
1174	*/
1175	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
1176	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgU32(volatile uint32_t RT_FAR *pu32, const uint32_t u32New, const uint32_t u32Old) RT_NOTHROW_PROTO;
1177	#else
1178	DECLINLINE(bool) ASMAtomicCmpXchgU32(volatile uint32_t RT_FAR *pu32, const uint32_t u32New, uint32_t u32Old) RT_NOTHROW_DEF
1179	{
1180	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1181	# if RT_INLINE_ASM_GNU_STYLE
1182	uint8_t u8Ret;
1183	__asm__ __volatile__("lock; cmpxchgl %3, %0\n\t"
1184	"setz %1\n\t"
1185	: "=m" (*pu32)
1186	, "=qm" (u8Ret)
1187	, "=a" (u32Old)
1188	: "r" (u32New)
1189	, "2" (u32Old)
1190	, "m" (*pu32)
1191	: "cc");
1192	return (bool)u8Ret;
1193
1194	# elif RT_INLINE_ASM_USES_INTRIN
1195	return (uint32_t)_InterlockedCompareExchange((long RT_FAR *)pu32, u32New, u32Old) == u32Old;
1196
1197	# else
1198	uint32_t u32Ret;
1199	__asm
1200	{
1201	# ifdef RT_ARCH_AMD64
1202	mov rdx, [pu32]
1203	# else
1204	mov edx, [pu32]
1205	# endif
1206	mov eax, [u32Old]
1207	mov ecx, [u32New]
1208	# ifdef RT_ARCH_AMD64
1209	lock cmpxchg [rdx], ecx
1210	# else
1211	lock cmpxchg [edx], ecx
1212	# endif
1213	setz al
1214	movzx eax, al
1215	mov [u32Ret], eax
1216	}
1217	return !!u32Ret;
1218	# endif
1219
1220	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
1221	union { uint32_t u; bool f; } fXchg;
1222	uint32_t u32Spill;
1223	uint32_t rcSpill;
1224	__asm__ __volatile__(".Ltry_again_ASMAtomicCmpXchgU32_%=:\n\t"
1225	RTASM_ARM_DMB_SY
1226	# if defined(RT_ARCH_ARM64)
1227	"ldaxr %w[uOld], %[pMem]\n\t"
1228	"cmp %w[uOld], %w[uCmp]\n\t"
1229	"bne 1f\n\t" /* stop here if not equal */
1230	"stlxr %w[rc], %w[uNew], %[pMem]\n\t"
1231	"cbnz %w[rc], .Ltry_again_ASMAtomicCmpXchgU32_%=\n\t"
1232	"mov %w[fXchg], #1\n\t"
1233	# else
1234	"ldrex %[uOld], %[pMem]\n\t"
1235	"teq %[uOld], %[uCmp]\n\t"
1236	"strexeq %[rc], %[uNew], %[pMem]\n\t"
1237	"bne 1f\n\t" /* stop here if not equal */
1238	"cmp %[rc], #0\n\t"
1239	"bne .Ltry_again_ASMAtomicCmpXchgU32_%=\n\t"
1240	"mov %[fXchg], #1\n\t"
1241	# endif
1242	"1:\n\t"
1243	: [pMem] "+Q" (*pu32)
1244	, [uOld] "=&r" (u32Spill)
1245	, [rc] "=&r" (rcSpill)
1246	, [fXchg] "=&r" (fXchg.u)
1247	: [uCmp] "r" (u32Old)
1248	, [uNew] "r" (u32New)
1249	, "[fXchg]" (0)
1250	RTASM_ARM_DMB_SY_COMMA_IN_REG
1251	: "cc");
1252	return fXchg.f;
1253
1254	# else
1255	# error "Port me"
1256	# endif
1257	}
1258	#endif
1259
1260
1261	/**
1262	* Atomically Compare and Exchange a signed 32-bit value, ordered.
1263	*
1264	* @returns true if xchg was done.
1265	* @returns false if xchg wasn't done.
1266	*
1267	* @param pi32 Pointer to the value to update.
1268	* @param i32New The new value to assigned to *pi32.
1269	* @param i32Old The old value to *pi32 compare with.
1270	*
1271	* @remarks x86: Requires a 486 or later.
1272	* @todo Rename ASMAtomicCmpWriteS32
1273	*/
1274	DECLINLINE(bool) ASMAtomicCmpXchgS32(volatile int32_t RT_FAR *pi32, const int32_t i32New, const int32_t i32Old) RT_NOTHROW_DEF
1275	{
1276	return ASMAtomicCmpXchgU32((volatile uint32_t RT_FAR *)pi32, (uint32_t)i32New, (uint32_t)i32Old);
1277	}
1278
1279
1280	/**
1281	* Atomically Compare and exchange an unsigned 64-bit value, ordered.
1282	*
1283	* @returns true if xchg was done.
1284	* @returns false if xchg wasn't done.
1285	*
1286	* @param pu64 Pointer to the 64-bit variable to update.
1287	* @param u64New The 64-bit value to assign to *pu64.
1288	* @param u64Old The value to compare with.
1289	*
1290	* @remarks x86: Requires a Pentium or later.
1291	* @todo Rename ASMAtomicCmpWriteU64
1292	*/
1293	#if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \
1294	\|\| RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
1295	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgU64(volatile uint64_t RT_FAR *pu64, const uint64_t u64New, const uint64_t u64Old) RT_NOTHROW_PROTO;
1296	#else
1297	DECLINLINE(bool) ASMAtomicCmpXchgU64(volatile uint64_t RT_FAR *pu64, uint64_t u64New, uint64_t u64Old) RT_NOTHROW_DEF
1298	{
1299	# if RT_INLINE_ASM_USES_INTRIN
1300	return (uint64_t)_InterlockedCompareExchange64((__int64 RT_FAR *)pu64, u64New, u64Old) == u64Old;
1301
1302	# elif defined(RT_ARCH_AMD64)
1303	# if RT_INLINE_ASM_GNU_STYLE
1304	uint8_t u8Ret;
1305	__asm__ __volatile__("lock; cmpxchgq %3, %0\n\t"
1306	"setz %1\n\t"
1307	: "=m" (*pu64)
1308	, "=qm" (u8Ret)
1309	, "=a" (u64Old)
1310	: "r" (u64New)
1311	, "2" (u64Old)
1312	, "m" (*pu64)
1313	: "cc");
1314	return (bool)u8Ret;
1315	# else
1316	bool fRet;
1317	__asm
1318	{
1319	mov rdx, [pu32]
1320	mov rax, [u64Old]
1321	mov rcx, [u64New]
1322	lock cmpxchg [rdx], rcx
1323	setz al
1324	mov [fRet], al
1325	}
1326	return fRet;
1327	# endif
1328
1329	# elif defined(RT_ARCH_X86)
1330	uint32_t u32Ret;
1331	# if RT_INLINE_ASM_GNU_STYLE
1332	# if defined(PIC) \|\| defined(__PIC__)
1333	uint32_t u32EBX = (uint32_t)u64New;
1334	uint32_t u32Spill;
1335	__asm__ __volatile__("xchgl %%ebx, %4\n\t"
1336	"lock; cmpxchg8b (%6)\n\t"
1337	"setz %%al\n\t"
1338	"movl %4, %%ebx\n\t"
1339	"movzbl %%al, %%eax\n\t"
1340	: "=a" (u32Ret)
1341	, "=d" (u32Spill)
1342	# if RT_GNUC_PREREQ(4, 3)
1343	, "+m" (*pu64)
1344	# else
1345	, "=m" (*pu64)
1346	# endif
1347	: "A" (u64Old)
1348	, "m" ( u32EBX )
1349	, "c" ( (uint32_t)(u64New >> 32) )
1350	, "S" (pu64)
1351	: "cc");
1352	# else /* !PIC */
1353	uint32_t u32Spill;
1354	__asm__ __volatile__("lock; cmpxchg8b %2\n\t"
1355	"setz %%al\n\t"
1356	"movzbl %%al, %%eax\n\t"
1357	: "=a" (u32Ret)
1358	, "=d" (u32Spill)
1359	, "+m" (*pu64)
1360	: "A" (u64Old)
1361	, "b" ( (uint32_t)u64New )
1362	, "c" ( (uint32_t)(u64New >> 32) )
1363	: "cc");
1364	# endif
1365	return (bool)u32Ret;
1366	# else
1367	__asm
1368	{
1369	mov ebx, dword ptr [u64New]
1370	mov ecx, dword ptr [u64New + 4]
1371	mov edi, [pu64]
1372	mov eax, dword ptr [u64Old]
1373	mov edx, dword ptr [u64Old + 4]
1374	lock cmpxchg8b [edi]
1375	setz al
1376	movzx eax, al
1377	mov dword ptr [u32Ret], eax
1378	}
1379	return !!u32Ret;
1380	# endif
1381
1382	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
1383	union { uint32_t u; bool f; } fXchg;
1384	uint64_t u64Spill;
1385	uint32_t rcSpill;
1386	__asm__ __volatile__(".Ltry_again_ASMAtomicCmpXchgU64_%=:\n\t"
1387	RTASM_ARM_DMB_SY
1388	# if defined(RT_ARCH_ARM64)
1389	"ldaxr %[uOld], %[pMem]\n\t"
1390	"cmp %[uOld], %[uCmp]\n\t"
1391	"bne 1f\n\t" /* stop here if not equal */
1392	"stlxr %w[rc], %[uNew], %[pMem]\n\t"
1393	"cbnz %w[rc], .Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
1394	"mov %w[fXchg], #1\n\t"
1395	# else
1396	"ldrexd %[uOld], %H[uOld], %[pMem]\n\t"
1397	"teq %[uOld], %[uCmp]\n\t"
1398	"teqeq %H[uOld], %H[uCmp]\n\t"
1399	"strexdeq %[rc], %[uNew], %H[uNew], %[pMem]\n\t"
1400	"bne 1f\n\t" /* stop here if not equal */
1401	"cmp %[rc], #0\n\t"
1402	"bne .Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
1403	"mov %[fXchg], #1\n\t"
1404	# endif
1405	"1:\n\t"
1406	: [pMem] "+Q" (*pu64)
1407	, [uOld] "=&r" (u64Spill)
1408	, [rc] "=&r" (rcSpill)
1409	, [fXchg] "=&r" (fXchg.u)
1410	: [uCmp] "r" (u64Old)
1411	, [uNew] "r" (u64New)
1412	, "[fXchg]" (0)
1413	RTASM_ARM_DMB_SY_COMMA_IN_REG
1414	: "cc");
1415	return fXchg.f;
1416
1417	# else
1418	# error "Port me"
1419	# endif
1420	}
1421	#endif
1422
1423
1424	/**
1425	* Atomically Compare and exchange a signed 64-bit value, ordered.
1426	*
1427	* @returns true if xchg was done.
1428	* @returns false if xchg wasn't done.
1429	*
1430	* @param pi64 Pointer to the 64-bit variable to update.
1431	* @param i64 The 64-bit value to assign to *pu64.
1432	* @param i64Old The value to compare with.
1433	*
1434	* @remarks x86: Requires a Pentium or later.
1435	* @todo Rename ASMAtomicCmpWriteS64
1436	*/
1437	DECLINLINE(bool) ASMAtomicCmpXchgS64(volatile int64_t RT_FAR *pi64, const int64_t i64, const int64_t i64Old) RT_NOTHROW_DEF
1438	{
1439	return ASMAtomicCmpXchgU64((volatile uint64_t RT_FAR *)pi64, (uint64_t)i64, (uint64_t)i64Old);
1440	}
1441
1442	#if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_ARM64) \|\| defined(DOXYGEN_RUNNING)
1443
1444	/** @def RTASM_HAVE_CMP_WRITE_U128
1445	* Indicates that we've got ASMAtomicCmpWriteU128(), ASMAtomicCmpWriteU128v2()
1446	* and ASMAtomicCmpWriteExU128() available. */
1447	# define RTASM_HAVE_CMP_WRITE_U128 1
1448
1449
1450	/**
1451	* Atomically compare and write an unsigned 128-bit value, ordered.
1452	*
1453	* @returns true if write was done.
1454	* @returns false if write wasn't done.
1455	*
1456	* @param pu128 Pointer to the 128-bit variable to update.
1457	* @param u64NewHi The high 64 bits of the value to assign to *pu128.
1458	* @param u64NewLo The low 64 bits of the value to assign to *pu128.
1459	* @param u64OldHi The high 64-bit of the value to compare with.
1460	* @param u64OldLo The low 64-bit of the value to compare with.
1461	*
1462	* @remarks AMD64: Not present in the earliest CPUs, so check CPUID.
1463	*/
1464	# if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN)
1465	DECLASM(bool) ASMAtomicCmpWriteU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo,
1466	const uint64_t u64OldHi, const uint64_t u64OldLo) RT_NOTHROW_PROTO;
1467	# else
1468	DECLINLINE(bool) ASMAtomicCmpWriteU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo,
1469	const uint64_t u64OldHi, const uint64_t u64OldLo) RT_NOTHROW_DEF
1470	{
1471	# if RT_INLINE_ASM_USES_INTRIN
1472	__int64 ai64Cmp[2];
1473	ai64Cmp[0] = u64OldLo;
1474	ai64Cmp[1] = u64OldHi;
1475	return _InterlockedCompareExchange128((__int64 volatile *)pu128, u64NewHi, u64NewLo, ai64Cmp) != 0;
1476
1477	# elif (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
1478	return __sync_bool_compare_and_swap(pu128, ((uint128_t)u64OldHi << 64) \| u64OldLo, ((uint128_t)u64NewHi << 64) \| u64NewLo);
1479
1480	# elif defined(RT_ARCH_AMD64)
1481	# if RT_INLINE_ASM_GNU_STYLE
1482	uint64_t u64Ret;
1483	uint64_t u64Spill;
1484	__asm__ __volatile__("lock; cmpxchg16b %2\n\t"
1485	"setz %%al\n\t"
1486	"movzbl %%al, %%eax\n\t"
1487	: "=a" (u64Ret)
1488	, "=d" (u64Spill)
1489	, "+m" (*pu128)
1490	: "a" (u64OldLo)
1491	, "d" (u64OldHi)
1492	, "b" (u64NewLo)
1493	, "c" (u64NewHi)
1494	: "cc");
1495
1496	return (bool)u64Ret;
1497	# else
1498	# error "Port me"
1499	# endif
1500	# else
1501	# error "Port me"
1502	# endif
1503	}
1504	# endif
1505
1506
1507	/**
1508	* Atomically compare and write an unsigned 128-bit value, ordered.
1509	*
1510	* @returns true if write was done.
1511	* @returns false if write wasn't done.
1512	*
1513	* @param pu128 Pointer to the 128-bit variable to update.
1514	* @param u128New The 128-bit value to assign to *pu128.
1515	* @param u128Old The value to compare with.
1516	*
1517	* @remarks AMD64: Not present in the earliest CPUs, so check CPUID.
1518	*/
1519	DECLINLINE(bool) ASMAtomicCmpWriteU128(volatile uint128_t *pu128, const uint128_t u128New, const uint128_t u128Old) RT_NOTHROW_DEF
1520	{
1521	# ifdef RT_COMPILER_WITH_128BIT_INT_TYPES
1522	# if (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
1523	return __sync_bool_compare_and_swap(pu128, u128Old, u128New);
1524	# else
1525	return ASMAtomicCmpWriteU128v2(pu128, (uint64_t)(u128New >> 64), (uint64_t)u128New,
1526	(uint64_t)(u128Old >> 64), (uint64_t)u128Old);
1527	# endif
1528	# else
1529	return ASMAtomicCmpWriteU128v2(pu128, u128New.Hi, u128New.Lo, u128Old.Hi, u128Old.Lo);
1530	# endif
1531	}
1532
1533
1534	/**
1535	* RTUINT128U wrapper for ASMAtomicCmpWriteU128.
1536	*/
1537	DECLINLINE(bool) ASMAtomicCmpWriteU128U(volatile RTUINT128U *pu128, const RTUINT128U u128New,
1538	const RTUINT128U u128Old) RT_NOTHROW_DEF
1539	{
1540	# if (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
1541	return ASMAtomicCmpWriteU128(&pu128->u, u128New.u, u128Old.u);
1542	# else
1543	return ASMAtomicCmpWriteU128v2(&pu128->u, u128New.s.Hi, u128New.s.Lo, u128Old.s.Hi, u128Old.s.Lo);
1544	# endif
1545	}
1546
1547	#endif /* RT_ARCH_AMD64 \|\| RT_ARCH_ARM64 */
1548
1549	/**
1550	* Atomically Compare and Exchange a pointer value, ordered.
1551	*
1552	* @returns true if xchg was done.
1553	* @returns false if xchg wasn't done.
1554	*
1555	* @param ppv Pointer to the value to update.
1556	* @param pvNew The new value to assigned to *ppv.
1557	* @param pvOld The old value to *ppv compare with.
1558	*
1559	* @remarks x86: Requires a 486 or later.
1560	* @todo Rename ASMAtomicCmpWritePtrVoid
1561	*/
1562	DECLINLINE(bool) ASMAtomicCmpXchgPtrVoid(void RT_FAR * volatile RT_FAR ppv, const void RT_FAR pvNew, const void RT_FAR *pvOld) RT_NOTHROW_DEF
1563	{
1564	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
1565	return ASMAtomicCmpXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )ppv, (uint32_t)pvNew, (uint32_t)pvOld);
1566	#elif ARCH_BITS == 64
1567	return ASMAtomicCmpXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )ppv, (uint64_t)pvNew, (uint64_t)pvOld);
1568	#else
1569	# error "ARCH_BITS is bogus"
1570	#endif
1571	}
1572
1573
1574	/**
1575	* Atomically Compare and Exchange a pointer value, ordered.
1576	*
1577	* @returns true if xchg was done.
1578	* @returns false if xchg wasn't done.
1579	*
1580	* @param ppv Pointer to the value to update.
1581	* @param pvNew The new value to assigned to *ppv.
1582	* @param pvOld The old value to *ppv compare with.
1583	*
1584	* @remarks This is relatively type safe on GCC platforms.
1585	* @remarks x86: Requires a 486 or later.
1586	* @todo Rename ASMAtomicCmpWritePtr
1587	*/
1588	#ifdef __GNUC__
1589	# define ASMAtomicCmpXchgPtr(ppv, pvNew, pvOld) \
1590	__extension__ \
1591	({\
1592	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
1593	__typeof__(*(ppv)) const pvNewTypeChecked = (pvNew); \
1594	__typeof__(*(ppv)) const pvOldTypeChecked = (pvOld); \
1595	bool fMacroRet = ASMAtomicCmpXchgPtrVoid((void * volatile *)ppvTypeChecked, \
1596	(void )pvNewTypeChecked, (void )pvOldTypeChecked); \
1597	fMacroRet; \
1598	})
1599	#else
1600	# define ASMAtomicCmpXchgPtr(ppv, pvNew, pvOld) \
1601	ASMAtomicCmpXchgPtrVoid((void RT_FAR * volatile RT_FAR )(ppv), (void RT_FAR )(pvNew), (void RT_FAR *)(pvOld))
1602	#endif
1603
1604
1605	/** @def ASMAtomicCmpXchgHandle
1606	* Atomically Compare and Exchange a typical IPRT handle value, ordered.
1607	*
1608	* @param ph Pointer to the value to update.
1609	* @param hNew The new value to assigned to *pu.
1610	* @param hOld The old value to *pu compare with.
1611	* @param fRc Where to store the result.
1612	*
1613	* @remarks This doesn't currently work for all handles (like RTFILE).
1614	* @remarks x86: Requires a 486 or later.
1615	* @todo Rename ASMAtomicCmpWriteHandle
1616	*/
1617	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
1618	# define ASMAtomicCmpXchgHandle(ph, hNew, hOld, fRc) \
1619	do { \
1620	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
1621	(fRc) = ASMAtomicCmpXchgU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)(hNew), (const uint32_t)(hOld)); \
1622	} while (0)
1623	#elif HC_ARCH_BITS == 64
1624	# define ASMAtomicCmpXchgHandle(ph, hNew, hOld, fRc) \
1625	do { \
1626	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
1627	(fRc) = ASMAtomicCmpXchgU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)(hNew), (const uint64_t)(hOld)); \
1628	} while (0)
1629	#else
1630	# error HC_ARCH_BITS
1631	#endif
1632
1633
1634	/** @def ASMAtomicCmpXchgSize
1635	* Atomically Compare and Exchange a value which size might differ
1636	* between platforms or compilers, ordered.
1637	*
1638	* @param pu Pointer to the value to update.
1639	* @param uNew The new value to assigned to *pu.
1640	* @param uOld The old value to *pu compare with.
1641	* @param fRc Where to store the result.
1642	*
1643	* @remarks x86: Requires a 486 or later.
1644	* @todo Rename ASMAtomicCmpWriteSize
1645	*/
1646	#define ASMAtomicCmpXchgSize(pu, uNew, uOld, fRc) \
1647	do { \
1648	switch (sizeof(*(pu))) { \
1649	case 4: (fRc) = ASMAtomicCmpXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew), (uint32_t)(uOld)); \
1650	break; \
1651	case 8: (fRc) = ASMAtomicCmpXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew), (uint64_t)(uOld)); \
1652	break; \
1653	default: AssertMsgFailed(("ASMAtomicCmpXchgSize: size %d is not supported\n", sizeof(*(pu)))); \
1654	(fRc) = false; \
1655	break; \
1656	} \
1657	} while (0)
1658
1659
1660	/**
1661	* Atomically Compare and Exchange an unsigned 8-bit value, additionally passes
1662	* back old value, ordered.
1663	*
1664	* @returns true if xchg was done.
1665	* @returns false if xchg wasn't done.
1666	*
1667	* @param pu8 Pointer to the value to update.
1668	* @param u8New The new value to assigned to *pu32.
1669	* @param u8Old The old value to *pu8 compare with.
1670	* @param pu8Old Pointer store the old value at.
1671	*
1672	* @remarks x86: Requires a 486 or later.
1673	*/
1674	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
1675	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgExU8(volatile uint8_t RT_FAR pu8, const uint8_t u8New, const uint8_t u8Old, uint8_t RT_FAR pu8Old) RT_NOTHROW_PROTO;
1676	#else
1677	DECLINLINE(bool) ASMAtomicCmpXchgExU8(volatile uint8_t RT_FAR pu8, const uint8_t u8New, const uint8_t u8Old, uint8_t RT_FAR pu8Old) RT_NOTHROW_DEF
1678	{
1679	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1680	# if RT_INLINE_ASM_GNU_STYLE
1681	uint8_t u8Ret;
1682	__asm__ __volatile__("lock; cmpxchgb %3, %0\n\t"
1683	"setz %1\n\t"
1684	: "=m" (*pu8)
1685	, "=qm" (u8Ret)
1686	, "=a" (*pu8Old)
1687	# if defined(RT_ARCH_X86)
1688	: "q" (u8New)
1689	# else
1690	: "r" (u8New)
1691	# endif
1692	, "a" (u8Old)
1693	, "m" (*pu8)
1694	: "cc");
1695	return (bool)u8Ret;
1696
1697	# elif RT_INLINE_ASM_USES_INTRIN
1698	return (pu8Old = _InterlockedCompareExchange8((char RT_FAR )pu8, u8New, u8Old)) == u8Old;
1699
1700	# else
1701	uint8_t u8Ret;
1702	__asm
1703	{
1704	# ifdef RT_ARCH_AMD64
1705	mov rdx, [pu8]
1706	# else
1707	mov edx, [pu8]
1708	# endif
1709	mov eax, [u8Old]
1710	mov ecx, [u8New]
1711	# ifdef RT_ARCH_AMD64
1712	lock cmpxchg [rdx], ecx
1713	mov rdx, [pu8Old]
1714	mov [rdx], eax
1715	# else
1716	lock cmpxchg [edx], ecx
1717	mov edx, [pu8Old]
1718	mov [edx], eax
1719	# endif
1720	setz al
1721	movzx eax, al
1722	mov [u8Ret], eax
1723	}
1724	return !!u8Ret;
1725	# endif
1726
1727	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
1728	union { uint8_t u; bool f; } fXchg;
1729	uint8_t u8ActualOld;
1730	uint8_t rcSpill;
1731	__asm__ __volatile__(".Ltry_again_ASMAtomicCmpXchgExU8_%=:\n\t"
1732	RTASM_ARM_DMB_SY
1733	# if defined(RT_ARCH_ARM64)
1734	"ldaxrb %w[uOld], %[pMem]\n\t"
1735	"cmp %w[uOld], %w[uCmp]\n\t"
1736	"bne 1f\n\t" /* stop here if not equal */
1737	"stlxrb %w[rc], %w[uNew], %[pMem]\n\t"
1738	"cbnz %w[rc], .Ltry_again_ASMAtomicCmpXchgExU8_%=\n\t"
1739	"mov %w[fXchg], #1\n\t"
1740	# else
1741	"ldrexb %[uOld], %[pMem]\n\t"
1742	"teq %[uOld], %[uCmp]\n\t"
1743	"strexbeq %[rc], %[uNew], %[pMem]\n\t"
1744	"bne 1f\n\t" /* stop here if not equal */
1745	"cmp %[rc], #0\n\t"
1746	"bne .Ltry_again_ASMAtomicCmpXchgExU8_%=\n\t"
1747	"mov %[fXchg], #1\n\t"
1748	# endif
1749	"1:\n\t"
1750	: [pMem] "+Q" (*pu8)
1751	, [uOld] "=&r" (u8ActualOld)
1752	, [rc] "=&r" (rcSpill)
1753	, [fXchg] "=&r" (fXchg.u)
1754	: [uCmp] "r" (u8Old)
1755	, [uNew] "r" (u8New)
1756	, "[fXchg]" (0)
1757	RTASM_ARM_DMB_SY_COMMA_IN_REG
1758	: "cc");
1759	*pu8Old = u8ActualOld;
1760	return fXchg.f;
1761
1762	# else
1763	# error "Port me"
1764	# endif
1765	}
1766	#endif
1767
1768
1769	/**
1770	* Atomically Compare and Exchange a signed 8-bit value, additionally
1771	* passes back old value, ordered.
1772	*
1773	* @returns true if xchg was done.
1774	* @returns false if xchg wasn't done.
1775	*
1776	* @param pi8 Pointer to the value to update.
1777	* @param i8New The new value to assigned to *pi8.
1778	* @param i8Old The old value to *pi8 compare with.
1779	* @param pi8Old Pointer store the old value at.
1780	*
1781	* @remarks x86: Requires a 486 or later.
1782	*/
1783	DECLINLINE(bool) ASMAtomicCmpXchgExS8(volatile int8_t RT_FAR pi8, const int8_t i8New, const int8_t i8Old, int8_t RT_FAR pi8Old) RT_NOTHROW_DEF
1784	{
1785	return ASMAtomicCmpXchgExU8((volatile uint8_t RT_FAR )pi8, (uint8_t)i8New, (uint8_t)i8Old, (uint8_t RT_FAR )pi8Old);
1786	}
1787
1788
1789	/**
1790	* Atomically Compare and Exchange an unsigned 16-bit value, additionally passes
1791	* back old value, ordered.
1792	*
1793	* @returns true if xchg was done.
1794	* @returns false if xchg wasn't done.
1795	*
1796	* @param pu16 Pointer to the value to update.
1797	* @param u16New The new value to assigned to *pu16.
1798	* @param u16Old The old value to *pu32 compare with.
1799	* @param pu16Old Pointer store the old value at.
1800	*
1801	* @remarks x86: Requires a 486 or later.
1802	*/
1803	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
1804	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgExU16(volatile uint16_t RT_FAR pu16, const uint16_t u16New, const uint16_t u16Old, uint16_t RT_FAR pu16Old) RT_NOTHROW_PROTO;
1805	#else
1806	DECLINLINE(bool) ASMAtomicCmpXchgExU16(volatile uint16_t RT_FAR pu16, const uint16_t u16New, const uint16_t u16Old, uint16_t RT_FAR pu16Old) RT_NOTHROW_DEF
1807	{
1808	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1809	# if RT_INLINE_ASM_GNU_STYLE
1810	uint8_t u8Ret;
1811	__asm__ __volatile__("lock; cmpxchgw %3, %0\n\t"
1812	"setz %1\n\t"
1813	: "=m" (*pu16)
1814	, "=qm" (u8Ret)
1815	, "=a" (*pu16Old)
1816	: "r" (u16New)
1817	, "a" (u16Old)
1818	, "m" (*pu16)
1819	: "cc");
1820	return (bool)u8Ret;
1821
1822	# elif RT_INLINE_ASM_USES_INTRIN
1823	return (pu16Old = _InterlockedCompareExchange16((short RT_FAR )pu16, u16New, u16Old)) == u16Old;
1824
1825	# else
1826	uint16_t u16Ret;
1827	__asm
1828	{
1829	# ifdef RT_ARCH_AMD64
1830	mov rdx, [pu16]
1831	# else
1832	mov edx, [pu16]
1833	# endif
1834	mov eax, [u16Old]
1835	mov ecx, [u16New]
1836	# ifdef RT_ARCH_AMD64
1837	lock cmpxchg [rdx], ecx
1838	mov rdx, [pu16Old]
1839	mov [rdx], eax
1840	# else
1841	lock cmpxchg [edx], ecx
1842	mov edx, [pu16Old]
1843	mov [edx], eax
1844	# endif
1845	setz al
1846	movzx eax, al
1847	mov [u16Ret], eax
1848	}
1849	return !!u16Ret;
1850	# endif
1851
1852	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
1853	union { uint16_t u; bool f; } fXchg;
1854	uint16_t u16ActualOld;
1855	uint16_t rcSpill;
1856	__asm__ __volatile__(".Ltry_again_ASMAtomicCmpXchgExU16_%=:\n\t"
1857	RTASM_ARM_DMB_SY
1858	# if defined(RT_ARCH_ARM64)
1859	"ldaxrh %w[uOld], %[pMem]\n\t"
1860	"cmp %w[uOld], %w[uCmp]\n\t"
1861	"bne 1f\n\t" /* stop here if not equal */
1862	"stlxrh %w[rc], %w[uNew], %[pMem]\n\t"
1863	"cbnz %w[rc], .Ltry_again_ASMAtomicCmpXchgExU16_%=\n\t"
1864	"mov %w[fXchg], #1\n\t"
1865	# else
1866	"ldrexh %[uOld], %[pMem]\n\t"
1867	"teq %[uOld], %[uCmp]\n\t"
1868	"strexheq %[rc], %[uNew], %[pMem]\n\t"
1869	"bne 1f\n\t" /* stop here if not equal */
1870	"cmp %[rc], #0\n\t"
1871	"bne .Ltry_again_ASMAtomicCmpXchgExU16_%=\n\t"
1872	"mov %[fXchg], #1\n\t"
1873	# endif
1874	"1:\n\t"
1875	: [pMem] "+Q" (*pu16)
1876	, [uOld] "=&r" (u16ActualOld)
1877	, [rc] "=&r" (rcSpill)
1878	, [fXchg] "=&r" (fXchg.u)
1879	: [uCmp] "r" (u16Old)
1880	, [uNew] "r" (u16New)
1881	, "[fXchg]" (0)
1882	RTASM_ARM_DMB_SY_COMMA_IN_REG
1883	: "cc");
1884	*pu16Old = u16ActualOld;
1885	return fXchg.f;
1886
1887	# else
1888	# error "Port me"
1889	# endif
1890	}
1891	#endif
1892
1893
1894	/**
1895	* Atomically Compare and Exchange a signed 16-bit value, additionally
1896	* passes back old value, ordered.
1897	*
1898	* @returns true if xchg was done.
1899	* @returns false if xchg wasn't done.
1900	*
1901	* @param pi16 Pointer to the value to update.
1902	* @param i16New The new value to assigned to *pi16.
1903	* @param i16Old The old value to *pi16 compare with.
1904	* @param pi16Old Pointer store the old value at.
1905	*
1906	* @remarks x86: Requires a 486 or later.
1907	*/
1908	DECLINLINE(bool) ASMAtomicCmpXchgExS16(volatile int16_t RT_FAR pi16, const int16_t i16New, const int16_t i16Old, int16_t RT_FAR pi16Old) RT_NOTHROW_DEF
1909	{
1910	return ASMAtomicCmpXchgExU16((volatile uint16_t RT_FAR )pi16, (uint16_t)i16New, (uint16_t)i16Old, (uint16_t RT_FAR )pi16Old);
1911	}
1912
1913
1914	/**
1915	* Atomically Compare and Exchange an unsigned 32-bit value, additionally
1916	* passes back old value, ordered.
1917	*
1918	* @returns true if xchg was done.
1919	* @returns false if xchg wasn't done.
1920	*
1921	* @param pu32 Pointer to the value to update.
1922	* @param u32New The new value to assigned to *pu32.
1923	* @param u32Old The old value to *pu32 compare with.
1924	* @param pu32Old Pointer store the old value at.
1925	*
1926	* @remarks x86: Requires a 486 or later.
1927	*/
1928	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
1929	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgExU32(volatile uint32_t RT_FAR pu32, const uint32_t u32New, const uint32_t u32Old, uint32_t RT_FAR pu32Old) RT_NOTHROW_PROTO;
1930	#else
1931	DECLINLINE(bool) ASMAtomicCmpXchgExU32(volatile uint32_t RT_FAR pu32, const uint32_t u32New, const uint32_t u32Old, uint32_t RT_FAR pu32Old) RT_NOTHROW_DEF
1932	{
1933	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1934	# if RT_INLINE_ASM_GNU_STYLE
1935	uint8_t u8Ret;
1936	__asm__ __volatile__("lock; cmpxchgl %3, %0\n\t"
1937	"setz %1\n\t"
1938	: "=m" (*pu32)
1939	, "=qm" (u8Ret)
1940	, "=a" (*pu32Old)
1941	: "r" (u32New)
1942	, "a" (u32Old)
1943	, "m" (*pu32)
1944	: "cc");
1945	return (bool)u8Ret;
1946
1947	# elif RT_INLINE_ASM_USES_INTRIN
1948	return (pu32Old = _InterlockedCompareExchange((long RT_FAR )pu32, u32New, u32Old)) == u32Old;
1949
1950	# else
1951	uint32_t u32Ret;
1952	__asm
1953	{
1954	# ifdef RT_ARCH_AMD64
1955	mov rdx, [pu32]
1956	# else
1957	mov edx, [pu32]
1958	# endif
1959	mov eax, [u32Old]
1960	mov ecx, [u32New]
1961	# ifdef RT_ARCH_AMD64
1962	lock cmpxchg [rdx], ecx
1963	mov rdx, [pu32Old]
1964	mov [rdx], eax
1965	# else
1966	lock cmpxchg [edx], ecx
1967	mov edx, [pu32Old]
1968	mov [edx], eax
1969	# endif
1970	setz al
1971	movzx eax, al
1972	mov [u32Ret], eax
1973	}
1974	return !!u32Ret;
1975	# endif
1976
1977	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
1978	union { uint32_t u; bool f; } fXchg;
1979	uint32_t u32ActualOld;
1980	uint32_t rcSpill;
1981	__asm__ __volatile__(".Ltry_again_ASMAtomicCmpXchgExU32_%=:\n\t"
1982	RTASM_ARM_DMB_SY
1983	# if defined(RT_ARCH_ARM64)
1984	"ldaxr %w[uOld], %[pMem]\n\t"
1985	"cmp %w[uOld], %w[uCmp]\n\t"
1986	"bne 1f\n\t" /* stop here if not equal */
1987	"stlxr %w[rc], %w[uNew], %[pMem]\n\t"
1988	"cbnz %w[rc], .Ltry_again_ASMAtomicCmpXchgExU32_%=\n\t"
1989	"mov %w[fXchg], #1\n\t"
1990	# else
1991	"ldrex %[uOld], %[pMem]\n\t"
1992	"teq %[uOld], %[uCmp]\n\t"
1993	"strexeq %[rc], %[uNew], %[pMem]\n\t"
1994	"bne 1f\n\t" /* stop here if not equal */
1995	"cmp %[rc], #0\n\t"
1996	"bne .Ltry_again_ASMAtomicCmpXchgExU32_%=\n\t"
1997	"mov %[fXchg], #1\n\t"
1998	# endif
1999	"1:\n\t"
2000	: [pMem] "+Q" (*pu32)
2001	, [uOld] "=&r" (u32ActualOld)
2002	, [rc] "=&r" (rcSpill)
2003	, [fXchg] "=&r" (fXchg.u)
2004	: [uCmp] "r" (u32Old)
2005	, [uNew] "r" (u32New)
2006	, "[fXchg]" (0)
2007	RTASM_ARM_DMB_SY_COMMA_IN_REG
2008	: "cc");
2009	*pu32Old = u32ActualOld;
2010	return fXchg.f;
2011
2012	# else
2013	# error "Port me"
2014	# endif
2015	}
2016	#endif
2017
2018
2019	/**
2020	* Atomically Compare and Exchange a signed 32-bit value, additionally
2021	* passes back old value, ordered.
2022	*
2023	* @returns true if xchg was done.
2024	* @returns false if xchg wasn't done.
2025	*
2026	* @param pi32 Pointer to the value to update.
2027	* @param i32New The new value to assigned to *pi32.
2028	* @param i32Old The old value to *pi32 compare with.
2029	* @param pi32Old Pointer store the old value at.
2030	*
2031	* @remarks x86: Requires a 486 or later.
2032	*/
2033	DECLINLINE(bool) ASMAtomicCmpXchgExS32(volatile int32_t RT_FAR pi32, const int32_t i32New, const int32_t i32Old, int32_t RT_FAR pi32Old) RT_NOTHROW_DEF
2034	{
2035	return ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR )pi32, (uint32_t)i32New, (uint32_t)i32Old, (uint32_t RT_FAR )pi32Old);
2036	}
2037
2038
2039	/**
2040	* Atomically Compare and exchange an unsigned 64-bit value, additionally
2041	* passing back old value, ordered.
2042	*
2043	* @returns true if xchg was done.
2044	* @returns false if xchg wasn't done.
2045	*
2046	* @param pu64 Pointer to the 64-bit variable to update.
2047	* @param u64New The 64-bit value to assign to *pu64.
2048	* @param u64Old The value to compare with.
2049	* @param pu64Old Pointer store the old value at.
2050	*
2051	* @remarks x86: Requires a Pentium or later.
2052	*/
2053	#if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \
2054	\|\| RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
2055	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgExU64(volatile uint64_t RT_FAR pu64, const uint64_t u64New, const uint64_t u64Old, uint64_t RT_FAR pu64Old) RT_NOTHROW_PROTO;
2056	#else
2057	DECLINLINE(bool) ASMAtomicCmpXchgExU64(volatile uint64_t RT_FAR pu64, const uint64_t u64New, const uint64_t u64Old, uint64_t RT_FAR pu64Old) RT_NOTHROW_DEF
2058	{
2059	# if RT_INLINE_ASM_USES_INTRIN
2060	return (pu64Old =_InterlockedCompareExchange64((__int64 RT_FAR )pu64, u64New, u64Old)) == u64Old;
2061
2062	# elif defined(RT_ARCH_AMD64)
2063	# if RT_INLINE_ASM_GNU_STYLE
2064	uint8_t u8Ret;
2065	__asm__ __volatile__("lock; cmpxchgq %3, %0\n\t"
2066	"setz %1\n\t"
2067	: "=m" (*pu64)
2068	, "=qm" (u8Ret)
2069	, "=a" (*pu64Old)
2070	: "r" (u64New)
2071	, "a" (u64Old)
2072	, "m" (*pu64)
2073	: "cc");
2074	return (bool)u8Ret;
2075	# else
2076	bool fRet;
2077	__asm
2078	{
2079	mov rdx, [pu32]
2080	mov rax, [u64Old]
2081	mov rcx, [u64New]
2082	lock cmpxchg [rdx], rcx
2083	mov rdx, [pu64Old]
2084	mov [rdx], rax
2085	setz al
2086	mov [fRet], al
2087	}
2088	return fRet;
2089	# endif
2090
2091	# elif defined(RT_ARCH_X86)
2092	# if RT_INLINE_ASM_GNU_STYLE
2093	uint64_t u64Ret;
2094	# if defined(PIC) \|\| defined(__PIC__)
2095	/* Note #1: This code uses a memory clobber description, because the clean
2096	solution with an output value for *pu64 makes gcc run out of
2097	registers. This will cause suboptimal code, and anyone with a
2098	better solution is welcome to improve this.
2099
2100	Note #2: We must prevent gcc from encoding the memory access, as it
2101	may go via the GOT if we're working on a global variable (like
2102	in the testcase). Thus we request a register (%3) and
2103	dereference it ourselves. */
2104	__asm__ __volatile__("xchgl %%ebx, %1\n\t"
2105	"lock; cmpxchg8b (%3)\n\t"
2106	"xchgl %%ebx, %1\n\t"
2107	: "=A" (u64Ret)
2108	: "DS" ((uint32_t)u64New)
2109	, "c" ((uint32_t)(u64New >> 32))
2110	, "r" (pu64) /* Do not use "m" here*/
2111	, "0" (u64Old)
2112	: "memory"
2113	, "cc" );
2114	# else /* !PIC */
2115	__asm__ __volatile__("lock; cmpxchg8b %4\n\t"
2116	: "=A" (u64Ret)
2117	, "=m" (*pu64)
2118	: "b" ((uint32_t)u64New)
2119	, "c" ((uint32_t)(u64New >> 32))
2120	, "m" (*pu64)
2121	, "0" (u64Old)
2122	: "cc");
2123	# endif
2124	*pu64Old = u64Ret;
2125	return u64Ret == u64Old;
2126	# else
2127	uint32_t u32Ret;
2128	__asm
2129	{
2130	mov ebx, dword ptr [u64New]
2131	mov ecx, dword ptr [u64New + 4]
2132	mov edi, [pu64]
2133	mov eax, dword ptr [u64Old]
2134	mov edx, dword ptr [u64Old + 4]
2135	lock cmpxchg8b [edi]
2136	mov ebx, [pu64Old]
2137	mov [ebx], eax
2138	setz al
2139	movzx eax, al
2140	add ebx, 4
2141	mov [ebx], edx
2142	mov dword ptr [u32Ret], eax
2143	}
2144	return !!u32Ret;
2145	# endif
2146
2147	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2148	union { uint32_t u; bool f; } fXchg;
2149	uint64_t u64ActualOld;
2150	uint32_t rcSpill;
2151	__asm__ __volatile__(".Ltry_again_ASMAtomicCmpXchgU64_%=:\n\t"
2152	RTASM_ARM_DMB_SY
2153	# if defined(RT_ARCH_ARM64)
2154	"ldaxr %[uOld], %[pMem]\n\t"
2155	"cmp %[uOld], %[uCmp]\n\t"
2156	"bne 1f\n\t" /* stop here if not equal */
2157	"stlxr %w[rc], %[uNew], %[pMem]\n\t"
2158	"cbnz %w[rc], .Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
2159	"mov %w[fXchg], #1\n\t"
2160	# else
2161	"ldrexd %[uOld], %H[uOld], %[pMem]\n\t"
2162	"teq %[uOld], %[uCmp]\n\t"
2163	"teqeq %H[uOld], %H[uCmp]\n\t"
2164	"strexdeq %[rc], %[uNew], %H[uNew], %[pMem]\n\t"
2165	"bne 1f\n\t" /* stop here if not equal */
2166	"cmp %[rc], #0\n\t"
2167	"bne .Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
2168	"mov %[fXchg], #1\n\t"
2169	# endif
2170	"1:\n\t"
2171	: [pMem] "+Q" (*pu64)
2172	, [uOld] "=&r" (u64ActualOld)
2173	, [rc] "=&r" (rcSpill)
2174	, [fXchg] "=&r" (fXchg.u)
2175	: [uCmp] "r" (u64Old)
2176	, [uNew] "r" (u64New)
2177	, "[fXchg]" (0)
2178	RTASM_ARM_DMB_SY_COMMA_IN_REG
2179	: "cc");
2180	*pu64Old = u64ActualOld;
2181	return fXchg.f;
2182
2183	# else
2184	# error "Port me"
2185	# endif
2186	}
2187	#endif
2188
2189
2190	/**
2191	* Atomically Compare and exchange a signed 64-bit value, additionally
2192	* passing back old value, ordered.
2193	*
2194	* @returns true if xchg was done.
2195	* @returns false if xchg wasn't done.
2196	*
2197	* @param pi64 Pointer to the 64-bit variable to update.
2198	* @param i64 The 64-bit value to assign to *pu64.
2199	* @param i64Old The value to compare with.
2200	* @param pi64Old Pointer store the old value at.
2201	*
2202	* @remarks x86: Requires a Pentium or later.
2203	*/
2204	DECLINLINE(bool) ASMAtomicCmpXchgExS64(volatile int64_t RT_FAR pi64, const int64_t i64, const int64_t i64Old, int64_t RT_FAR pi64Old) RT_NOTHROW_DEF
2205	{
2206	return ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR )pi64, (uint64_t)i64, (uint64_t)i64Old, (uint64_t RT_FAR )pi64Old);
2207	}
2208
2209	#if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_ARM64) \|\| defined(DOXYGEN_RUNNING)
2210
2211	/** @def RTASM_HAVE_CMP_XCHG_U128
2212	* Indicates that we've got ASMAtomicCmpSwapU128(), ASMAtomicCmpSwapU128v2()
2213	* and ASMAtomicCmpSwapExU128() available. */
2214	# define RTASM_HAVE_CMP_XCHG_U128 1
2215
2216
2217	/**
2218	* Atomically compare and exchange an unsigned 128-bit value, ordered.
2219	*
2220	* @returns true if exchange was done.
2221	* @returns false if exchange wasn't done.
2222	*
2223	* @param pu128 Pointer to the 128-bit variable to update.
2224	* @param u64NewHi The high 64 bits of the value to assign to *pu128.
2225	* @param u64NewLo The low 64 bits of the value to assign to *pu128.
2226	* @param u64OldHi The high 64-bit of the value to compare with.
2227	* @param u64OldLo The low 64-bit of the value to compare with.
2228	* @param pu128Old Where to return the old value.
2229	*
2230	* @remarks AMD64: Not present in the earliest CPUs, so check CPUID.
2231	*/
2232	# if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN)
2233	DECLASM(bool) ASMAtomicCmpXchgU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo,
2234	const uint64_t u64OldHi, const uint64_t u64OldLo, uint128_t *pu128Old) RT_NOTHROW_PROTO;
2235	# else
2236	DECLINLINE(bool) ASMAtomicCmpXchgU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo,
2237	const uint64_t u64OldHi, const uint64_t u64OldLo, uint128_t *pu128Old) RT_NOTHROW_DEF
2238	{
2239	# if RT_INLINE_ASM_USES_INTRIN
2240	pu128Old->Hi = u64OldHi;
2241	pu128Old->Lo = u64OldLo;
2242	AssertCompileMemberOffset(uint128_t, Lo, 0);
2243	return _InterlockedCompareExchange128((__int64 volatile )pu128, u64NewHi, u64NewLo, (__int64 )&pu128Old->Lo) != 0;
2244
2245	# elif (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
2246	uint128_t const uCmp = ((uint128_t)u64OldHi << 64) \| u64OldLo;
2247	uint128_t const uOld = __sync_val_compare_and_swap(pu128, uCmp, ((uint128_t)u64NewHi << 64) \| u64NewLo);
2248	*pu128Old = uOld;
2249	return uCmp == uOld;
2250
2251	# elif defined(RT_ARCH_AMD64)
2252	# if RT_INLINE_ASM_GNU_STYLE
2253	uint8_t bRet;
2254	uint64_t u64RetHi, u64RetLo;
2255	__asm__ __volatile__("lock; cmpxchg16b %3\n\t"
2256	"setz %b0\n\t"
2257	: "=r" (bRet)
2258	, "=a" (u64RetLo)
2259	, "=d" (u64RetHi)
2260	, "+m" (*pu128)
2261	: "a" (u64OldLo)
2262	, "d" (u64OldHi)
2263	, "b" (u64NewLo)
2264	, "c" (u64NewHi)
2265	: "cc");
2266	*pu128Old = ((uint128_t)u64RetHi << 64) \| u64RetLo;
2267	return (bool)bRet;
2268	# else
2269	# error "Port me"
2270	# endif
2271	# else
2272	# error "Port me"
2273	# endif
2274	}
2275	# endif
2276
2277
2278	/**
2279	* Atomically compare and exchange an unsigned 128-bit value, ordered.
2280	*
2281	* @returns true if exchange was done.
2282	* @returns false if exchange wasn't done.
2283	*
2284	* @param pu128 Pointer to the 128-bit variable to update.
2285	* @param u128New The 128-bit value to assign to *pu128.
2286	* @param u128Old The value to compare with.
2287	* @param pu128Old Where to return the old value.
2288	*
2289	* @remarks AMD64: Not present in the earliest CPUs, so check CPUID.
2290	*/
2291	DECLINLINE(bool) ASMAtomicCmpXchgU128(volatile uint128_t *pu128, const uint128_t u128New,
2292	const uint128_t u128Old, uint128_t *pu128Old) RT_NOTHROW_DEF
2293	{
2294	# ifdef RT_COMPILER_WITH_128BIT_INT_TYPES
2295	# if (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
2296	uint128_t const uSwapped = __sync_val_compare_and_swap(pu128, u128Old, u128New);
2297	*pu128Old = uSwapped;
2298	return uSwapped == u128Old;
2299	# else
2300	return ASMAtomicCmpXchgU128v2(pu128, (uint64_t)(u128New >> 64), (uint64_t)u128New,
2301	(uint64_t)(u128Old >> 64), (uint64_t)u128Old, pu128Old);
2302	# endif
2303	# else
2304	return ASMAtomicCmpXchgU128v2(pu128, u128New.Hi, u128New.Lo, u128Old.Hi, u128Old.Lo, pu128Old);
2305	# endif
2306	}
2307
2308
2309	/**
2310	* RTUINT128U wrapper for ASMAtomicCmpXchgU128.
2311	*/
2312	DECLINLINE(bool) ASMAtomicCmpXchgU128U(volatile RTUINT128U *pu128, const RTUINT128U u128New,
2313	const RTUINT128U u128Old, PRTUINT128U pu128Old) RT_NOTHROW_DEF
2314	{
2315	# if (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
2316	return ASMAtomicCmpXchgU128(&pu128->u, u128New.u, u128Old.u, &pu128Old->u);
2317	# else
2318	return ASMAtomicCmpXchgU128v2(&pu128->u, u128New.s.Hi, u128New.s.Lo, u128Old.s.Hi, u128Old.s.Lo, &pu128Old->u);
2319	# endif
2320	}
2321
2322	#endif /* RT_ARCH_AMD64 \|\| RT_ARCH_ARM64 */
2323
2324
2325
2326	/** @def ASMAtomicCmpXchgExHandle
2327	* Atomically Compare and Exchange a typical IPRT handle value, ordered.
2328	*
2329	* @param ph Pointer to the value to update.
2330	* @param hNew The new value to assigned to *pu.
2331	* @param hOld The old value to *pu compare with.
2332	* @param fRc Where to store the result.
2333	* @param phOldVal Pointer to where to store the old value.
2334	*
2335	* @remarks This doesn't currently work for all handles (like RTFILE).
2336	*/
2337	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
2338	# define ASMAtomicCmpXchgExHandle(ph, hNew, hOld, fRc, phOldVal) \
2339	do { \
2340	AssertCompile(sizeof(*ph) == sizeof(uint32_t)); \
2341	AssertCompile(sizeof(*phOldVal) == sizeof(uint32_t)); \
2342	(fRc) = ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR )(ph), (uint32_t)(hNew), (uint32_t)(hOld), (uint32_t RT_FAR )(phOldVal)); \
2343	} while (0)
2344	#elif HC_ARCH_BITS == 64
2345	# define ASMAtomicCmpXchgExHandle(ph, hNew, hOld, fRc, phOldVal) \
2346	do { \
2347	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
2348	AssertCompile(sizeof(*(phOldVal)) == sizeof(uint64_t)); \
2349	(fRc) = ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR )(ph), (uint64_t)(hNew), (uint64_t)(hOld), (uint64_t RT_FAR )(phOldVal)); \
2350	} while (0)
2351	#else
2352	# error HC_ARCH_BITS
2353	#endif
2354
2355
2356	/** @def ASMAtomicCmpXchgExSize
2357	* Atomically Compare and Exchange a value which size might differ
2358	* between platforms or compilers. Additionally passes back old value.
2359	*
2360	* @param pu Pointer to the value to update.
2361	* @param uNew The new value to assigned to *pu.
2362	* @param uOld The old value to *pu compare with.
2363	* @param fRc Where to store the result.
2364	* @param puOldVal Pointer to where to store the old value.
2365	*
2366	* @remarks x86: Requires a 486 or later.
2367	*/
2368	#define ASMAtomicCmpXchgExSize(pu, uNew, uOld, fRc, puOldVal) \
2369	do { \
2370	switch (sizeof(*(pu))) { \
2371	case 4: (fRc) = ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew), (uint32_t)(uOld), (uint32_t RT_FAR *)(uOldVal)); \
2372	break; \
2373	case 8: (fRc) = ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew), (uint64_t)(uOld), (uint64_t RT_FAR *)(uOldVal)); \
2374	break; \
2375	default: AssertMsgFailed(("ASMAtomicCmpXchgSize: size %d is not supported\n", sizeof(*(pu)))); \
2376	(fRc) = false; \
2377	(uOldVal) = 0; \
2378	break; \
2379	} \
2380	} while (0)
2381
2382
2383	/**
2384	* Atomically Compare and Exchange a pointer value, additionally
2385	* passing back old value, ordered.
2386	*
2387	* @returns true if xchg was done.
2388	* @returns false if xchg wasn't done.
2389	*
2390	* @param ppv Pointer to the value to update.
2391	* @param pvNew The new value to assigned to *ppv.
2392	* @param pvOld The old value to *ppv compare with.
2393	* @param ppvOld Pointer store the old value at.
2394	*
2395	* @remarks x86: Requires a 486 or later.
2396	*/
2397	DECLINLINE(bool) ASMAtomicCmpXchgExPtrVoid(void RT_FAR * volatile RT_FAR ppv, const void RT_FAR pvNew, const void RT_FAR *pvOld,
2398	void RT_FAR * RT_FAR *ppvOld) RT_NOTHROW_DEF
2399	{
2400	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
2401	return ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR )(void RT_FAR )ppv, (uint32_t)pvNew, (uint32_t)pvOld, (uint32_t RT_FAR *)ppvOld);
2402	#elif ARCH_BITS == 64
2403	return ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR )(void RT_FAR )ppv, (uint64_t)pvNew, (uint64_t)pvOld, (uint64_t RT_FAR *)ppvOld);
2404	#else
2405	# error "ARCH_BITS is bogus"
2406	#endif
2407	}
2408
2409
2410	/**
2411	* Atomically Compare and Exchange a pointer value, additionally
2412	* passing back old value, ordered.
2413	*
2414	* @returns true if xchg was done.
2415	* @returns false if xchg wasn't done.
2416	*
2417	* @param ppv Pointer to the value to update.
2418	* @param pvNew The new value to assigned to *ppv.
2419	* @param pvOld The old value to *ppv compare with.
2420	* @param ppvOld Pointer store the old value at.
2421	*
2422	* @remarks This is relatively type safe on GCC platforms.
2423	* @remarks x86: Requires a 486 or later.
2424	*/
2425	#ifdef __GNUC__
2426	# define ASMAtomicCmpXchgExPtr(ppv, pvNew, pvOld, ppvOld) \
2427	__extension__ \
2428	({\
2429	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
2430	__typeof__(*(ppv)) const pvNewTypeChecked = (pvNew); \
2431	__typeof__(*(ppv)) const pvOldTypeChecked = (pvOld); \
2432	__typeof__((ppv)) const ppvOldTypeChecked = (ppvOld); \
2433	bool fMacroRet = ASMAtomicCmpXchgExPtrVoid((void * volatile *)ppvTypeChecked, \
2434	(void )pvNewTypeChecked, (void )pvOldTypeChecked, \
2435	(void **)ppvOldTypeChecked); \
2436	fMacroRet; \
2437	})
2438	#else
2439	# define ASMAtomicCmpXchgExPtr(ppv, pvNew, pvOld, ppvOld) \
2440	ASMAtomicCmpXchgExPtrVoid((void RT_FAR * volatile RT_FAR )(ppv), (void RT_FAR )(pvNew), (void RT_FAR )(pvOld), (void RT_FAR RT_FAR *)(ppvOld))
2441	#endif
2442
2443
2444	/**
2445	* Virtualization unfriendly serializing instruction, always exits.
2446	*/
2447	#if (RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN) \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
2448	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSerializeInstructionCpuId(void) RT_NOTHROW_PROTO;
2449	#else
2450	DECLINLINE(void) ASMSerializeInstructionCpuId(void) RT_NOTHROW_DEF
2451	{
2452	# if RT_INLINE_ASM_GNU_STYLE
2453	RTCCUINTREG xAX = 0;
2454	# ifdef RT_ARCH_AMD64
2455	__asm__ __volatile__ ("cpuid"
2456	: "=a" (xAX)
2457	: "0" (xAX)
2458	: "rbx", "rcx", "rdx", "memory");
2459	# elif (defined(PIC) \|\| defined(__PIC__)) && defined(__i386__)
2460	__asm__ __volatile__ ("push %%ebx\n\t"
2461	"cpuid\n\t"
2462	"pop %%ebx\n\t"
2463	: "=a" (xAX)
2464	: "0" (xAX)
2465	: "ecx", "edx", "memory");
2466	# else
2467	__asm__ __volatile__ ("cpuid"
2468	: "=a" (xAX)
2469	: "0" (xAX)
2470	: "ebx", "ecx", "edx", "memory");
2471	# endif
2472
2473	# elif RT_INLINE_ASM_USES_INTRIN
2474	int aInfo[4];
2475	_ReadWriteBarrier();
2476	__cpuid(aInfo, 0);
2477
2478	# else
2479	__asm
2480	{
2481	push ebx
2482	xor eax, eax
2483	cpuid
2484	pop ebx
2485	}
2486	# endif
2487	}
2488	#endif
2489
2490	/**
2491	* Virtualization friendly serializing instruction, though more expensive.
2492	*/
2493	#if RT_INLINE_ASM_EXTERNAL \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
2494	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSerializeInstructionIRet(void) RT_NOTHROW_PROTO;
2495	#else
2496	DECLINLINE(void) ASMSerializeInstructionIRet(void) RT_NOTHROW_DEF
2497	{
2498	# if RT_INLINE_ASM_GNU_STYLE
2499	# ifdef RT_ARCH_AMD64
2500	__asm__ __volatile__ ("movq %%rsp,%%r10\n\t"
2501	"subq $128, %%rsp\n\t" /redzone/
2502	"mov %%ss, %%eax\n\t"
2503	"pushq %%rax\n\t"
2504	"pushq %%r10\n\t"
2505	"pushfq\n\t"
2506	"movl %%cs, %%eax\n\t"
2507	"pushq %%rax\n\t"
2508	"leaq 1f(%%rip), %%rax\n\t"
2509	"pushq %%rax\n\t"
2510	"iretq\n\t"
2511	"1:\n\t"
2512	::: "rax", "r10", "memory", "cc");
2513	# else
2514	__asm__ __volatile__ ("pushfl\n\t"
2515	"pushl %%cs\n\t"
2516	"pushl $1f\n\t"
2517	"iretl\n\t"
2518	"1:\n\t"
2519	::: "memory");
2520	# endif
2521
2522	# else
2523	__asm
2524	{
2525	pushfd
2526	push cs
2527	push la_ret
2528	iretd
2529	la_ret:
2530	}
2531	# endif
2532	}
2533	#endif
2534
2535	/**
2536	* Virtualization friendlier serializing instruction, may still cause exits.
2537	*/
2538	#if (RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2008) \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
2539	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSerializeInstructionRdTscp(void) RT_NOTHROW_PROTO;
2540	#else
2541	DECLINLINE(void) ASMSerializeInstructionRdTscp(void) RT_NOTHROW_DEF
2542	{
2543	# if RT_INLINE_ASM_GNU_STYLE
2544	/* rdtscp is not supported by ancient linux build VM of course :-( */
2545	# ifdef RT_ARCH_AMD64
2546	/__asm__ __volatile__("rdtscp\n\t" ::: "rax", "rdx, "rcx"); /
2547	__asm__ __volatile__(".byte 0x0f,0x01,0xf9\n\t" ::: "rax", "rdx", "rcx", "memory");
2548	# else
2549	/__asm__ __volatile__("rdtscp\n\t" ::: "eax", "edx, "ecx"); /
2550	__asm__ __volatile__(".byte 0x0f,0x01,0xf9\n\t" ::: "eax", "edx", "ecx", "memory");
2551	# endif
2552	# else
2553	# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2008
2554	uint32_t uIgnore;
2555	_ReadWriteBarrier();
2556	(void)__rdtscp(&uIgnore);
2557	(void)uIgnore;
2558	# else
2559	__asm
2560	{
2561	rdtscp
2562	}
2563	# endif
2564	# endif
2565	}
2566	#endif
2567
2568
2569	/**
2570	* Serialize Instruction (both data store and instruction flush).
2571	*/
2572	#if (defined(RT_ARCH_X86) && ARCH_BITS == 16) \|\| defined(IN_GUEST)
2573	# define ASMSerializeInstruction() ASMSerializeInstructionIRet()
2574	#elif defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)
2575	# define ASMSerializeInstruction() ASMSerializeInstructionCpuId()
2576	#elif defined(RT_ARCH_SPARC64)
2577	RTDECL(void) ASMSerializeInstruction(void) RT_NOTHROW_PROTO;
2578	#elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2579	DECLINLINE(void) ASMSerializeInstruction(void) RT_NOTHROW_DEF
2580	{
2581	__asm__ __volatile__ (RTASM_ARM_DSB_SY :: RTASM_ARM_DSB_SY_IN_REG :);
2582	}
2583	#else
2584	# error "Port me"
2585	#endif
2586
2587
2588	/**
2589	* Memory fence, waits for any pending writes and reads to complete.
2590	* @note No implicit compiler barrier (which is probably stupid).
2591	*/
2592	DECLINLINE(void) ASMMemoryFence(void) RT_NOTHROW_DEF
2593	{
2594	#if defined(RT_ARCH_AMD64) \|\| (defined(RT_ARCH_X86) && !defined(RT_WITH_OLD_CPU_SUPPORT))
2595	# if RT_INLINE_ASM_GNU_STYLE
2596	__asm__ __volatile__ (".byte 0x0f,0xae,0xf0\n\t");
2597	# elif RT_INLINE_ASM_USES_INTRIN
2598	_mm_mfence();
2599	# else
2600	__asm
2601	{
2602	_emit 0x0f
2603	_emit 0xae
2604	_emit 0xf0
2605	}
2606	# endif
2607	#elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2608	__asm__ __volatile__ (RTASM_ARM_DMB_SY :: RTASM_ARM_DMB_SY_IN_REG :);
2609	#elif ARCH_BITS == 16
2610	uint16_t volatile u16;
2611	ASMAtomicXchgU16(&u16, 0);
2612	#else
2613	uint32_t volatile u32;
2614	ASMAtomicXchgU32(&u32, 0);
2615	#endif
2616	}
2617
2618
2619	/**
2620	* Write fence, waits for any pending writes to complete.
2621	* @note No implicit compiler barrier (which is probably stupid).
2622	*/
2623	DECLINLINE(void) ASMWriteFence(void) RT_NOTHROW_DEF
2624	{
2625	#if defined(RT_ARCH_AMD64) \|\| (defined(RT_ARCH_X86) && !defined(RT_WITH_OLD_CPU_SUPPORT))
2626	# if RT_INLINE_ASM_GNU_STYLE
2627	__asm__ __volatile__ (".byte 0x0f,0xae,0xf8\n\t");
2628	# elif RT_INLINE_ASM_USES_INTRIN
2629	_mm_sfence();
2630	# else
2631	__asm
2632	{
2633	_emit 0x0f
2634	_emit 0xae
2635	_emit 0xf8
2636	}
2637	# endif
2638	#elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2639	__asm__ __volatile__ (RTASM_ARM_DMB_ST :: RTASM_ARM_DMB_ST_IN_REG :);
2640	#else
2641	ASMMemoryFence();
2642	#endif
2643	}
2644
2645
2646	/**
2647	* Read fence, waits for any pending reads to complete.
2648	* @note No implicit compiler barrier (which is probably stupid).
2649	*/
2650	DECLINLINE(void) ASMReadFence(void) RT_NOTHROW_DEF
2651	{
2652	#if defined(RT_ARCH_AMD64) \|\| (defined(RT_ARCH_X86) && !defined(RT_WITH_OLD_CPU_SUPPORT))
2653	# if RT_INLINE_ASM_GNU_STYLE
2654	__asm__ __volatile__ (".byte 0x0f,0xae,0xe8\n\t");
2655	# elif RT_INLINE_ASM_USES_INTRIN
2656	_mm_lfence();
2657	# else
2658	__asm
2659	{
2660	_emit 0x0f
2661	_emit 0xae
2662	_emit 0xe8
2663	}
2664	# endif
2665	#elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2666	__asm__ __volatile__ (RTASM_ARM_DMB_LD :: RTASM_ARM_DMB_LD_IN_REG :);
2667	#else
2668	ASMMemoryFence();
2669	#endif
2670	}
2671
2672
2673	/**
2674	* Atomically reads an unsigned 8-bit value, ordered.
2675	*
2676	* @returns Current *pu8 value
2677	* @param pu8 Pointer to the 8-bit variable to read.
2678	*/
2679	DECLINLINE(uint8_t) ASMAtomicReadU8(volatile uint8_t RT_FAR *pu8) RT_NOTHROW_DEF
2680	{
2681	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2682	uint32_t u32;
2683	__asm__ __volatile__(".Lstart_ASMAtomicReadU8_%=:\n\t"
2684	RTASM_ARM_DMB_SY
2685	# if defined(RT_ARCH_ARM64)
2686	"ldxrb %w[uDst], %[pMem]\n\t"
2687	# else
2688	"ldrexb %[uDst], %[pMem]\n\t"
2689	# endif
2690	: [uDst] "=&r" (u32)
2691	: [pMem] "Q" (*pu8)
2692	RTASM_ARM_DMB_SY_COMMA_IN_REG);
2693	return (uint8_t)u32;
2694	#else
2695	ASMMemoryFence();
2696	return pu8; / byte reads are atomic on x86 */
2697	#endif
2698	}
2699
2700
2701	/**
2702	* Atomically reads an unsigned 8-bit value, unordered.
2703	*
2704	* @returns Current *pu8 value
2705	* @param pu8 Pointer to the 8-bit variable to read.
2706	*/
2707	DECLINLINE(uint8_t) ASMAtomicUoReadU8(volatile uint8_t RT_FAR *pu8) RT_NOTHROW_DEF
2708	{
2709	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2710	uint32_t u32;
2711	__asm__ __volatile__(".Lstart_ASMAtomicUoReadU8_%=:\n\t"
2712	# if defined(RT_ARCH_ARM64)
2713	"ldxrb %w[uDst], %[pMem]\n\t"
2714	# else
2715	"ldrexb %[uDst], %[pMem]\n\t"
2716	# endif
2717	: [uDst] "=&r" (u32)
2718	: [pMem] "Q" (*pu8));
2719	return (uint8_t)u32;
2720	#else
2721	return pu8; / byte reads are atomic on x86 */
2722	#endif
2723	}
2724
2725
2726	/**
2727	* Atomically reads a signed 8-bit value, ordered.
2728	*
2729	* @returns Current *pi8 value
2730	* @param pi8 Pointer to the 8-bit variable to read.
2731	*/
2732	DECLINLINE(int8_t) ASMAtomicReadS8(volatile int8_t RT_FAR *pi8) RT_NOTHROW_DEF
2733	{
2734	ASMMemoryFence();
2735	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2736	int32_t i32;
2737	__asm__ __volatile__(".Lstart_ASMAtomicReadS8_%=:\n\t"
2738	RTASM_ARM_DMB_SY
2739	# if defined(RT_ARCH_ARM64)
2740	"ldxrb %w[iDst], %[pMem]\n\t"
2741	# else
2742	"ldrexb %[iDst], %[pMem]\n\t"
2743	# endif
2744	: [iDst] "=&r" (i32)
2745	: [pMem] "Q" (*pi8)
2746	RTASM_ARM_DMB_SY_COMMA_IN_REG);
2747	return (int8_t)i32;
2748	#else
2749	return pi8; / byte reads are atomic on x86 */
2750	#endif
2751	}
2752
2753
2754	/**
2755	* Atomically reads a signed 8-bit value, unordered.
2756	*
2757	* @returns Current *pi8 value
2758	* @param pi8 Pointer to the 8-bit variable to read.
2759	*/
2760	DECLINLINE(int8_t) ASMAtomicUoReadS8(volatile int8_t RT_FAR *pi8) RT_NOTHROW_DEF
2761	{
2762	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2763	int32_t i32;
2764	__asm__ __volatile__(".Lstart_ASMAtomicUoReadS8_%=:\n\t"
2765	# if defined(RT_ARCH_ARM64)
2766	"ldxrb %w[iDst], %[pMem]\n\t"
2767	# else
2768	"ldrexb %[iDst], %[pMem]\n\t"
2769	# endif
2770	: [iDst] "=&r" (i32)
2771	: [pMem] "Q" (*pi8));
2772	return (int8_t)i32;
2773	#else
2774	return pi8; / byte reads are atomic on x86 */
2775	#endif
2776	}
2777
2778
2779	/**
2780	* Atomically reads an unsigned 16-bit value, ordered.
2781	*
2782	* @returns Current *pu16 value
2783	* @param pu16 Pointer to the 16-bit variable to read.
2784	*/
2785	DECLINLINE(uint16_t) ASMAtomicReadU16(volatile uint16_t RT_FAR *pu16) RT_NOTHROW_DEF
2786	{
2787	Assert(!((uintptr_t)pu16 & 1));
2788	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2789	uint32_t u32;
2790	__asm__ __volatile__(".Lstart_ASMAtomicReadU16_%=:\n\t"
2791	RTASM_ARM_DMB_SY
2792	# if defined(RT_ARCH_ARM64)
2793	"ldxrh %w[uDst], %[pMem]\n\t"
2794	# else
2795	"ldrexh %[uDst], %[pMem]\n\t"
2796	# endif
2797	: [uDst] "=&r" (u32)
2798	: [pMem] "Q" (*pu16)
2799	RTASM_ARM_DMB_SY_COMMA_IN_REG);
2800	return (uint16_t)u32;
2801	#else
2802	ASMMemoryFence();
2803	return *pu16;
2804	#endif
2805	}
2806
2807
2808	/**
2809	* Atomically reads an unsigned 16-bit value, unordered.
2810	*
2811	* @returns Current *pu16 value
2812	* @param pu16 Pointer to the 16-bit variable to read.
2813	*/
2814	DECLINLINE(uint16_t) ASMAtomicUoReadU16(volatile uint16_t RT_FAR *pu16) RT_NOTHROW_DEF
2815	{
2816	Assert(!((uintptr_t)pu16 & 1));
2817	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2818	uint32_t u32;
2819	__asm__ __volatile__(".Lstart_ASMAtomicUoReadU16_%=:\n\t"
2820	# if defined(RT_ARCH_ARM64)
2821	"ldxrh %w[uDst], %[pMem]\n\t"
2822	# else
2823	"ldrexh %[uDst], %[pMem]\n\t"
2824	# endif
2825	: [uDst] "=&r" (u32)
2826	: [pMem] "Q" (*pu16));
2827	return (uint16_t)u32;
2828	#else
2829	return *pu16;
2830	#endif
2831	}
2832
2833
2834	/**
2835	* Atomically reads a signed 16-bit value, ordered.
2836	*
2837	* @returns Current *pi16 value
2838	* @param pi16 Pointer to the 16-bit variable to read.
2839	*/
2840	DECLINLINE(int16_t) ASMAtomicReadS16(volatile int16_t RT_FAR *pi16) RT_NOTHROW_DEF
2841	{
2842	Assert(!((uintptr_t)pi16 & 1));
2843	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2844	int32_t i32;
2845	__asm__ __volatile__(".Lstart_ASMAtomicReadS16_%=:\n\t"
2846	RTASM_ARM_DMB_SY
2847	# if defined(RT_ARCH_ARM64)
2848	"ldxrh %w[iDst], %[pMem]\n\t"
2849	# else
2850	"ldrexh %[iDst], %[pMem]\n\t"
2851	# endif
2852	: [iDst] "=&r" (i32)
2853	: [pMem] "Q" (*pi16)
2854	RTASM_ARM_DMB_SY_COMMA_IN_REG);
2855	return (int16_t)i32;
2856	#else
2857	ASMMemoryFence();
2858	return *pi16;
2859	#endif
2860	}
2861
2862
2863	/**
2864	* Atomically reads a signed 16-bit value, unordered.
2865	*
2866	* @returns Current *pi16 value
2867	* @param pi16 Pointer to the 16-bit variable to read.
2868	*/
2869	DECLINLINE(int16_t) ASMAtomicUoReadS16(volatile int16_t RT_FAR *pi16) RT_NOTHROW_DEF
2870	{
2871	Assert(!((uintptr_t)pi16 & 1));
2872	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2873	int32_t i32;
2874	__asm__ __volatile__(".Lstart_ASMAtomicUoReadS16_%=:\n\t"
2875	# if defined(RT_ARCH_ARM64)
2876	"ldxrh %w[iDst], %[pMem]\n\t"
2877	# else
2878	"ldrexh %[iDst], %[pMem]\n\t"
2879	# endif
2880	: [iDst] "=&r" (i32)
2881	: [pMem] "Q" (*pi16));
2882	return (int16_t)i32;
2883	#else
2884	return *pi16;
2885	#endif
2886	}
2887
2888
2889	/**
2890	* Atomically reads an unsigned 32-bit value, ordered.
2891	*
2892	* @returns Current *pu32 value
2893	* @param pu32 Pointer to the 32-bit variable to read.
2894	*/
2895	DECLINLINE(uint32_t) ASMAtomicReadU32(volatile uint32_t RT_FAR *pu32) RT_NOTHROW_DEF
2896	{
2897	Assert(!((uintptr_t)pu32 & 3));
2898	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2899	uint32_t u32;
2900	__asm__ __volatile__(".Lstart_ASMAtomicReadU32_%=:\n\t"
2901	RTASM_ARM_DMB_SY
2902	# if defined(RT_ARCH_ARM64)
2903	"ldxr %w[uDst], %[pMem]\n\t"
2904	# else
2905	"ldrex %[uDst], %[pMem]\n\t"
2906	# endif
2907	: [uDst] "=&r" (u32)
2908	: [pMem] "Q" (*pu32)
2909	RTASM_ARM_DMB_SY_COMMA_IN_REG);
2910	return u32;
2911	#else
2912	ASMMemoryFence();
2913	# if ARCH_BITS == 16
2914	AssertFailed(); /** @todo 16-bit */
2915	# endif
2916	return *pu32;
2917	#endif
2918	}
2919
2920
2921	/**
2922	* Atomically reads an unsigned 32-bit value, unordered.
2923	*
2924	* @returns Current *pu32 value
2925	* @param pu32 Pointer to the 32-bit variable to read.
2926	*/
2927	DECLINLINE(uint32_t) ASMAtomicUoReadU32(volatile uint32_t RT_FAR *pu32) RT_NOTHROW_DEF
2928	{
2929	Assert(!((uintptr_t)pu32 & 3));
2930	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2931	uint32_t u32;
2932	__asm__ __volatile__(".Lstart_ASMAtomicUoReadU32_%=:\n\t"
2933	# if defined(RT_ARCH_ARM64)
2934	"ldxr %w[uDst], %[pMem]\n\t"
2935	# else
2936	"ldrex %[uDst], %[pMem]\n\t"
2937	# endif
2938	: [uDst] "=&r" (u32)
2939	: [pMem] "Q" (*pu32));
2940	return u32;
2941	#else
2942	# if ARCH_BITS == 16
2943	AssertFailed(); /** @todo 16-bit */
2944	# endif
2945	return *pu32;
2946	#endif
2947	}
2948
2949
2950	/**
2951	* Atomically reads a signed 32-bit value, ordered.
2952	*
2953	* @returns Current *pi32 value
2954	* @param pi32 Pointer to the 32-bit variable to read.
2955	*/
2956	DECLINLINE(int32_t) ASMAtomicReadS32(volatile int32_t RT_FAR *pi32) RT_NOTHROW_DEF
2957	{
2958	Assert(!((uintptr_t)pi32 & 3));
2959	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2960	int32_t i32;
2961	__asm__ __volatile__(".Lstart_ASMAtomicReadS32_%=:\n\t"
2962	RTASM_ARM_DMB_SY
2963	# if defined(RT_ARCH_ARM64)
2964	"ldxr %w[iDst], %[pMem]\n\t"
2965	# else
2966	"ldrex %[iDst], %[pMem]\n\t"
2967	# endif
2968	: [iDst] "=&r" (i32)
2969	: [pMem] "Q" (*pi32)
2970	RTASM_ARM_DMB_SY_COMMA_IN_REG);
2971	return i32;
2972	#else
2973	ASMMemoryFence();
2974	# if ARCH_BITS == 16
2975	AssertFailed(); /** @todo 16-bit */
2976	# endif
2977	return *pi32;
2978	#endif
2979	}
2980
2981
2982	/**
2983	* Atomically reads a signed 32-bit value, unordered.
2984	*
2985	* @returns Current *pi32 value
2986	* @param pi32 Pointer to the 32-bit variable to read.
2987	*/
2988	DECLINLINE(int32_t) ASMAtomicUoReadS32(volatile int32_t RT_FAR *pi32) RT_NOTHROW_DEF
2989	{
2990	Assert(!((uintptr_t)pi32 & 3));
2991	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2992	int32_t i32;
2993	__asm__ __volatile__(".Lstart_ASMAtomicUoReadS32_%=:\n\t"
2994	# if defined(RT_ARCH_ARM64)
2995	"ldxr %w[iDst], %[pMem]\n\t"
2996	# else
2997	"ldrex %[iDst], %[pMem]\n\t"
2998	# endif
2999	: [iDst] "=&r" (i32)
3000	: [pMem] "Q" (*pi32));
3001	return i32;
3002
3003	#else
3004	# if ARCH_BITS == 16
3005	AssertFailed(); /** @todo 16-bit */
3006	# endif
3007	return *pi32;
3008	#endif
3009	}
3010
3011
3012	/**
3013	* Atomically reads an unsigned 64-bit value, ordered.
3014	*
3015	* @returns Current *pu64 value
3016	* @param pu64 Pointer to the 64-bit variable to read.
3017	* The memory pointed to must be writable.
3018	*
3019	* @remarks This may fault if the memory is read-only!
3020	* @remarks x86: Requires a Pentium or later.
3021	*/
3022	#if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !defined(RT_ARCH_AMD64)) \
3023	\|\| RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
3024	RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_PROTO;
3025	#else
3026	DECLINLINE(uint64_t) ASMAtomicReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_DEF
3027	{
3028	uint64_t u64;
3029	# ifdef RT_ARCH_AMD64
3030	Assert(!((uintptr_t)pu64 & 7));
3031	/*# if RT_INLINE_ASM_GNU_STYLE
3032	__asm__ __volatile__( "mfence\n\t"
3033	"movq %1, %0\n\t"
3034	: "=r" (u64)
3035	: "m" (*pu64));
3036	# else
3037	__asm
3038	{
3039	mfence
3040	mov rdx, [pu64]
3041	mov rax, [rdx]
3042	mov [u64], rax
3043	}
3044	# endif*/
3045	ASMMemoryFence();
3046	u64 = *pu64;
3047
3048	# elif defined(RT_ARCH_X86)
3049	# if RT_INLINE_ASM_GNU_STYLE
3050	# if defined(PIC) \|\| defined(__PIC__)
3051	uint32_t u32EBX = 0;
3052	Assert(!((uintptr_t)pu64 & 7));
3053	__asm__ __volatile__("xchgl %%ebx, %3\n\t"
3054	"lock; cmpxchg8b (%5)\n\t"
3055	"movl %3, %%ebx\n\t"
3056	: "=A" (u64)
3057	# if RT_GNUC_PREREQ(4, 3)
3058	, "+m" (*pu64)
3059	# else
3060	, "=m" (*pu64)
3061	# endif
3062	: "0" (0ULL)
3063	, "m" (u32EBX)
3064	, "c" (0)
3065	, "S" (pu64)
3066	: "cc");
3067	# else /* !PIC */
3068	__asm__ __volatile__("lock; cmpxchg8b %1\n\t"
3069	: "=A" (u64)
3070	, "+m" (*pu64)
3071	: "0" (0ULL)
3072	, "b" (0)
3073	, "c" (0)
3074	: "cc");
3075	# endif
3076	# else
3077	Assert(!((uintptr_t)pu64 & 7));
3078	__asm
3079	{
3080	xor eax, eax
3081	xor edx, edx
3082	mov edi, pu64
3083	xor ecx, ecx
3084	xor ebx, ebx
3085	lock cmpxchg8b [edi]
3086	mov dword ptr [u64], eax
3087	mov dword ptr [u64 + 4], edx
3088	}
3089	# endif
3090
3091	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3092	Assert(!((uintptr_t)pu64 & 7));
3093	__asm__ __volatile__(".Lstart_ASMAtomicReadU64_%=:\n\t"
3094	RTASM_ARM_DMB_SY
3095	# if defined(RT_ARCH_ARM64)
3096	"ldxr %[uDst], %[pMem]\n\t"
3097	# else
3098	"ldrexd %[uDst], %H[uDst], %[pMem]\n\t"
3099	# endif
3100	: [uDst] "=&r" (u64)
3101	: [pMem] "Q" (*pu64)
3102	RTASM_ARM_DMB_SY_COMMA_IN_REG);
3103
3104	# else
3105	# error "Port me"
3106	# endif
3107	return u64;
3108	}
3109	#endif
3110
3111
3112	/**
3113	* Atomically reads an unsigned 64-bit value, unordered.
3114	*
3115	* @returns Current *pu64 value
3116	* @param pu64 Pointer to the 64-bit variable to read.
3117	* The memory pointed to must be writable.
3118	*
3119	* @remarks This may fault if the memory is read-only!
3120	* @remarks x86: Requires a Pentium or later.
3121	*/
3122	#if !defined(RT_ARCH_AMD64) \
3123	&& ( (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \
3124	\|\| RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC)
3125	RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicUoReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_PROTO;
3126	#else
3127	DECLINLINE(uint64_t) ASMAtomicUoReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_DEF
3128	{
3129	uint64_t u64;
3130	# ifdef RT_ARCH_AMD64
3131	Assert(!((uintptr_t)pu64 & 7));
3132	/*# if RT_INLINE_ASM_GNU_STYLE
3133	Assert(!((uintptr_t)pu64 & 7));
3134	__asm__ __volatile__("movq %1, %0\n\t"
3135	: "=r" (u64)
3136	: "m" (*pu64));
3137	# else
3138	__asm
3139	{
3140	mov rdx, [pu64]
3141	mov rax, [rdx]
3142	mov [u64], rax
3143	}
3144	# endif */
3145	u64 = *pu64;
3146
3147	# elif defined(RT_ARCH_X86)
3148	# if RT_INLINE_ASM_GNU_STYLE
3149	# if defined(PIC) \|\| defined(__PIC__)
3150	uint32_t u32EBX = 0;
3151	uint32_t u32Spill;
3152	Assert(!((uintptr_t)pu64 & 7));
3153	__asm__ __volatile__("xor %%eax,%%eax\n\t"
3154	"xor %%ecx,%%ecx\n\t"
3155	"xor %%edx,%%edx\n\t"
3156	"xchgl %%ebx, %3\n\t"
3157	"lock; cmpxchg8b (%4)\n\t"
3158	"movl %3, %%ebx\n\t"
3159	: "=A" (u64)
3160	# if RT_GNUC_PREREQ(4, 3)
3161	, "+m" (*pu64)
3162	# else
3163	, "=m" (*pu64)
3164	# endif
3165	, "=c" (u32Spill)
3166	: "m" (u32EBX)
3167	, "S" (pu64)
3168	: "cc");
3169	# else /* !PIC */
3170	__asm__ __volatile__("lock; cmpxchg8b %1\n\t"
3171	: "=A" (u64)
3172	, "+m" (*pu64)
3173	: "0" (0ULL)
3174	, "b" (0)
3175	, "c" (0)
3176	: "cc");
3177	# endif
3178	# else
3179	Assert(!((uintptr_t)pu64 & 7));
3180	__asm
3181	{
3182	xor eax, eax
3183	xor edx, edx
3184	mov edi, pu64
3185	xor ecx, ecx
3186	xor ebx, ebx
3187	lock cmpxchg8b [edi]
3188	mov dword ptr [u64], eax
3189	mov dword ptr [u64 + 4], edx
3190	}
3191	# endif
3192
3193	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3194	Assert(!((uintptr_t)pu64 & 7));
3195	__asm__ __volatile__(".Lstart_ASMAtomicUoReadU64_%=:\n\t"
3196	# if defined(RT_ARCH_ARM64)
3197	"ldxr %[uDst], %[pMem]\n\t"
3198	# else
3199	"ldrexd %[uDst], %H[uDst], %[pMem]\n\t"
3200	# endif
3201	: [uDst] "=&r" (u64)
3202	: [pMem] "Q" (*pu64));
3203
3204	# else
3205	# error "Port me"
3206	# endif
3207	return u64;
3208	}
3209	#endif
3210
3211
3212	/**
3213	* Atomically reads a signed 64-bit value, ordered.
3214	*
3215	* @returns Current *pi64 value
3216	* @param pi64 Pointer to the 64-bit variable to read.
3217	* The memory pointed to must be writable.
3218	*
3219	* @remarks This may fault if the memory is read-only!
3220	* @remarks x86: Requires a Pentium or later.
3221	*/
3222	DECLINLINE(int64_t) ASMAtomicReadS64(volatile int64_t RT_FAR *pi64) RT_NOTHROW_DEF
3223	{
3224	return (int64_t)ASMAtomicReadU64((volatile uint64_t RT_FAR *)pi64);
3225	}
3226
3227
3228	/**
3229	* Atomically reads a signed 64-bit value, unordered.
3230	*
3231	* @returns Current *pi64 value
3232	* @param pi64 Pointer to the 64-bit variable to read.
3233	* The memory pointed to must be writable.
3234	*
3235	* @remarks This will fault if the memory is read-only!
3236	* @remarks x86: Requires a Pentium or later.
3237	*/
3238	DECLINLINE(int64_t) ASMAtomicUoReadS64(volatile int64_t RT_FAR *pi64) RT_NOTHROW_DEF
3239	{
3240	return (int64_t)ASMAtomicUoReadU64((volatile uint64_t RT_FAR *)pi64);
3241	}
3242
3243
3244	/**
3245	* Atomically reads a size_t value, ordered.
3246	*
3247	* @returns Current *pcb value
3248	* @param pcb Pointer to the size_t variable to read.
3249	*/
3250	DECLINLINE(size_t) ASMAtomicReadZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF
3251	{
3252	#if ARCH_BITS == 64
3253	return ASMAtomicReadU64((uint64_t volatile RT_FAR *)pcb);
3254	#elif ARCH_BITS == 32
3255	return ASMAtomicReadU32((uint32_t volatile RT_FAR *)pcb);
3256	#elif ARCH_BITS == 16
3257	AssertCompileSize(size_t, 2);
3258	return ASMAtomicReadU16((uint16_t volatile RT_FAR *)pcb);
3259	#else
3260	# error "Unsupported ARCH_BITS value"
3261	#endif
3262	}
3263
3264
3265	/**
3266	* Atomically reads a size_t value, unordered.
3267	*
3268	* @returns Current *pcb value
3269	* @param pcb Pointer to the size_t variable to read.
3270	*/
3271	DECLINLINE(size_t) ASMAtomicUoReadZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF
3272	{
3273	#if ARCH_BITS == 64 \|\| ARCH_BITS == 16
3274	return ASMAtomicUoReadU64((uint64_t volatile RT_FAR *)pcb);
3275	#elif ARCH_BITS == 32
3276	return ASMAtomicUoReadU32((uint32_t volatile RT_FAR *)pcb);
3277	#elif ARCH_BITS == 16
3278	AssertCompileSize(size_t, 2);
3279	return ASMAtomicUoReadU16((uint16_t volatile RT_FAR *)pcb);
3280	#else
3281	# error "Unsupported ARCH_BITS value"
3282	#endif
3283	}
3284
3285
3286	/**
3287	* Atomically reads a pointer value, ordered.
3288	*
3289	* @returns Current *pv value
3290	* @param ppv Pointer to the pointer variable to read.
3291	*
3292	* @remarks Please use ASMAtomicReadPtrT, it provides better type safety and
3293	* requires less typing (no casts).
3294	*/
3295	DECLINLINE(void RT_FAR ) ASMAtomicReadPtr(void RT_FAR volatile RT_FAR *ppv) RT_NOTHROW_DEF
3296	{
3297	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
3298	return (void RT_FAR )ASMAtomicReadU32((volatile uint32_t RT_FAR )(void RT_FAR *)ppv);
3299	#elif ARCH_BITS == 64
3300	return (void RT_FAR )ASMAtomicReadU64((volatile uint64_t RT_FAR )(void RT_FAR *)ppv);
3301	#else
3302	# error "ARCH_BITS is bogus"
3303	#endif
3304	}
3305
3306	/**
3307	* Convenience macro for avoiding the annoying casting with ASMAtomicReadPtr.
3308	*
3309	* @returns Current *pv value
3310	* @param ppv Pointer to the pointer variable to read.
3311	* @param Type The type of *ppv, sans volatile.
3312	*/
3313	#ifdef __GNUC__ /* 8.2.0 requires -Wno-ignored-qualifiers */
3314	# define ASMAtomicReadPtrT(ppv, Type) \
3315	__extension__ \
3316	({\
3317	__typeof__((ppv)) volatile ppvTypeChecked = (ppv); \
3318	Type pvTypeChecked = (__typeof__((ppv))) ASMAtomicReadPtr((void volatile *)ppvTypeChecked); \
3319	pvTypeChecked; \
3320	})
3321	#else
3322	# define ASMAtomicReadPtrT(ppv, Type) \
3323	(Type)ASMAtomicReadPtr((void RT_FAR * volatile RT_FAR *)(ppv))
3324	#endif
3325
3326
3327	/**
3328	* Atomically reads a pointer value, unordered.
3329	*
3330	* @returns Current *pv value
3331	* @param ppv Pointer to the pointer variable to read.
3332	*
3333	* @remarks Please use ASMAtomicUoReadPtrT, it provides better type safety and
3334	* requires less typing (no casts).
3335	*/
3336	DECLINLINE(void RT_FAR ) ASMAtomicUoReadPtr(void RT_FAR volatile RT_FAR *ppv) RT_NOTHROW_DEF
3337	{
3338	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
3339	return (void RT_FAR )ASMAtomicUoReadU32((volatile uint32_t RT_FAR )(void RT_FAR *)ppv);
3340	#elif ARCH_BITS == 64
3341	return (void RT_FAR )ASMAtomicUoReadU64((volatile uint64_t RT_FAR )(void RT_FAR *)ppv);
3342	#else
3343	# error "ARCH_BITS is bogus"
3344	#endif
3345	}
3346
3347
3348	/**
3349	* Convenience macro for avoiding the annoying casting with ASMAtomicUoReadPtr.
3350	*
3351	* @returns Current *pv value
3352	* @param ppv Pointer to the pointer variable to read.
3353	* @param Type The type of *ppv, sans volatile.
3354	*/
3355	#ifdef __GNUC__ /* 8.2.0 requires -Wno-ignored-qualifiers */
3356	# define ASMAtomicUoReadPtrT(ppv, Type) \
3357	__extension__ \
3358	({\
3359	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
3360	Type pvTypeChecked = (__typeof__((ppv))) ASMAtomicUoReadPtr((void volatile *)ppvTypeChecked); \
3361	pvTypeChecked; \
3362	})
3363	#else
3364	# define ASMAtomicUoReadPtrT(ppv, Type) \
3365	(Type)ASMAtomicUoReadPtr((void RT_FAR * volatile RT_FAR *)(ppv))
3366	#endif
3367
3368
3369	/**
3370	* Atomically reads a boolean value, ordered.
3371	*
3372	* @returns Current *pf value
3373	* @param pf Pointer to the boolean variable to read.
3374	*/
3375	DECLINLINE(bool) ASMAtomicReadBool(volatile bool RT_FAR *pf) RT_NOTHROW_DEF
3376	{
3377	ASMMemoryFence();
3378	return pf; / byte reads are atomic on x86 */
3379	}
3380
3381
3382	/**
3383	* Atomically reads a boolean value, unordered.
3384	*
3385	* @returns Current *pf value
3386	* @param pf Pointer to the boolean variable to read.
3387	*/
3388	DECLINLINE(bool) ASMAtomicUoReadBool(volatile bool RT_FAR *pf) RT_NOTHROW_DEF
3389	{
3390	return pf; / byte reads are atomic on x86 */
3391	}
3392
3393
3394	/**
3395	* Atomically read a typical IPRT handle value, ordered.
3396	*
3397	* @param ph Pointer to the handle variable to read.
3398	* @param phRes Where to store the result.
3399	*
3400	* @remarks This doesn't currently work for all handles (like RTFILE).
3401	*/
3402	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
3403	# define ASMAtomicReadHandle(ph, phRes) \
3404	do { \
3405	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
3406	AssertCompile(sizeof(*(phRes)) == sizeof(uint32_t)); \
3407	(uint32_t RT_FAR )(phRes) = ASMAtomicReadU32((uint32_t volatile RT_FAR *)(ph)); \
3408	} while (0)
3409	#elif HC_ARCH_BITS == 64
3410	# define ASMAtomicReadHandle(ph, phRes) \
3411	do { \
3412	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
3413	AssertCompile(sizeof(*(phRes)) == sizeof(uint64_t)); \
3414	(uint64_t RT_FAR )(phRes) = ASMAtomicReadU64((uint64_t volatile RT_FAR *)(ph)); \
3415	} while (0)
3416	#else
3417	# error HC_ARCH_BITS
3418	#endif
3419
3420
3421	/**
3422	* Atomically read a typical IPRT handle value, unordered.
3423	*
3424	* @param ph Pointer to the handle variable to read.
3425	* @param phRes Where to store the result.
3426	*
3427	* @remarks This doesn't currently work for all handles (like RTFILE).
3428	*/
3429	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
3430	# define ASMAtomicUoReadHandle(ph, phRes) \
3431	do { \
3432	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
3433	AssertCompile(sizeof(*(phRes)) == sizeof(uint32_t)); \
3434	(uint32_t RT_FAR )(phRes) = ASMAtomicUoReadU32((uint32_t volatile RT_FAR *)(ph)); \
3435	} while (0)
3436	#elif HC_ARCH_BITS == 64
3437	# define ASMAtomicUoReadHandle(ph, phRes) \
3438	do { \
3439	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
3440	AssertCompile(sizeof(*(phRes)) == sizeof(uint64_t)); \
3441	(uint64_t RT_FAR )(phRes) = ASMAtomicUoReadU64((uint64_t volatile RT_FAR *)(ph)); \
3442	} while (0)
3443	#else
3444	# error HC_ARCH_BITS
3445	#endif
3446
3447
3448	/**
3449	* Atomically read a value which size might differ
3450	* between platforms or compilers, ordered.
3451	*
3452	* @param pu Pointer to the variable to read.
3453	* @param puRes Where to store the result.
3454	*/
3455	#define ASMAtomicReadSize(pu, puRes) \
3456	do { \
3457	switch (sizeof(*(pu))) { \
3458	case 1: (uint8_t RT_FAR )(puRes) = ASMAtomicReadU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu)); break; \
3459	case 2: (uint16_t RT_FAR )(puRes) = ASMAtomicReadU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu)); break; \
3460	case 4: (uint32_t RT_FAR )(puRes) = ASMAtomicReadU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu)); break; \
3461	case 8: (uint64_t RT_FAR )(puRes) = ASMAtomicReadU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu)); break; \
3462	default: AssertMsgFailed(("ASMAtomicReadSize: size %d is not supported\n", sizeof(*(pu)))); \
3463	} \
3464	} while (0)
3465
3466
3467	/**
3468	* Atomically read a value which size might differ
3469	* between platforms or compilers, unordered.
3470	*
3471	* @param pu Pointer to the variable to read.
3472	* @param puRes Where to store the result.
3473	*/
3474	#define ASMAtomicUoReadSize(pu, puRes) \
3475	do { \
3476	switch (sizeof(*(pu))) { \
3477	case 1: (uint8_t RT_FAR )(puRes) = ASMAtomicUoReadU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu)); break; \
3478	case 2: (uint16_t RT_FAR )(puRes) = ASMAtomicUoReadU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu)); break; \
3479	case 4: (uint32_t RT_FAR )(puRes) = ASMAtomicUoReadU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu)); break; \
3480	case 8: (uint64_t RT_FAR )(puRes) = ASMAtomicUoReadU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu)); break; \
3481	default: AssertMsgFailed(("ASMAtomicReadSize: size %d is not supported\n", sizeof(*(pu)))); \
3482	} \
3483	} while (0)
3484
3485
3486	/**
3487	* Atomically writes an unsigned 8-bit value, ordered.
3488	*
3489	* @param pu8 Pointer to the 8-bit variable.
3490	* @param u8 The 8-bit value to assign to *pu8.
3491	*/
3492	DECLINLINE(void) ASMAtomicWriteU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_DEF
3493	{
3494	/** @todo Any possible ARM32/ARM64 optimizations here? */
3495	ASMAtomicXchgU8(pu8, u8);
3496	}
3497
3498
3499	/**
3500	* Atomically writes an unsigned 8-bit value, unordered.
3501	*
3502	* @param pu8 Pointer to the 8-bit variable.
3503	* @param u8 The 8-bit value to assign to *pu8.
3504	*/
3505	DECLINLINE(void) ASMAtomicUoWriteU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_DEF
3506	{
3507	/** @todo Any possible ARM32/ARM64 improvements here? */
3508	pu8 = u8; / byte writes are atomic on x86 */
3509	}
3510
3511
3512	/**
3513	* Atomically writes a signed 8-bit value, ordered.
3514	*
3515	* @param pi8 Pointer to the 8-bit variable to read.
3516	* @param i8 The 8-bit value to assign to *pi8.
3517	*/
3518	DECLINLINE(void) ASMAtomicWriteS8(volatile int8_t RT_FAR *pi8, int8_t i8) RT_NOTHROW_DEF
3519	{
3520	/** @todo Any possible ARM32/ARM64 optimizations here? */
3521	ASMAtomicXchgS8(pi8, i8);
3522	}
3523
3524
3525	/**
3526	* Atomically writes a signed 8-bit value, unordered.
3527	*
3528	* @param pi8 Pointer to the 8-bit variable to write.
3529	* @param i8 The 8-bit value to assign to *pi8.
3530	*/
3531	DECLINLINE(void) ASMAtomicUoWriteS8(volatile int8_t RT_FAR *pi8, int8_t i8) RT_NOTHROW_DEF
3532	{
3533	pi8 = i8; / byte writes are atomic on x86 */
3534	}
3535
3536
3537	/**
3538	* Atomically writes an unsigned 16-bit value, ordered.
3539	*
3540	* @param pu16 Pointer to the 16-bit variable to write.
3541	* @param u16 The 16-bit value to assign to *pu16.
3542	*/
3543	DECLINLINE(void) ASMAtomicWriteU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_DEF
3544	{
3545	/** @todo Any possible ARM32/ARM64 optimizations here? */
3546	ASMAtomicXchgU16(pu16, u16);
3547	}
3548
3549
3550	/**
3551	* Atomically writes an unsigned 16-bit value, unordered.
3552	*
3553	* @param pu16 Pointer to the 16-bit variable to write.
3554	* @param u16 The 16-bit value to assign to *pu16.
3555	*/
3556	DECLINLINE(void) ASMAtomicUoWriteU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_DEF
3557	{
3558	Assert(!((uintptr_t)pu16 & 1));
3559	*pu16 = u16;
3560	}
3561
3562
3563	/**
3564	* Atomically writes a signed 16-bit value, ordered.
3565	*
3566	* @param pi16 Pointer to the 16-bit variable to write.
3567	* @param i16 The 16-bit value to assign to *pi16.
3568	*/
3569	DECLINLINE(void) ASMAtomicWriteS16(volatile int16_t RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF
3570	{
3571	/** @todo Any possible ARM32/ARM64 optimizations here? */
3572	ASMAtomicXchgS16(pi16, i16);
3573	}
3574
3575
3576	/**
3577	* Atomically writes a signed 16-bit value, unordered.
3578	*
3579	* @param pi16 Pointer to the 16-bit variable to write.
3580	* @param i16 The 16-bit value to assign to *pi16.
3581	*/
3582	DECLINLINE(void) ASMAtomicUoWriteS16(volatile int16_t RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF
3583	{
3584	Assert(!((uintptr_t)pi16 & 1));
3585	*pi16 = i16;
3586	}
3587
3588
3589	/**
3590	* Atomically writes an unsigned 32-bit value, ordered.
3591	*
3592	* @param pu32 Pointer to the 32-bit variable to write.
3593	* @param u32 The 32-bit value to assign to *pu32.
3594	*/
3595	DECLINLINE(void) ASMAtomicWriteU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
3596	{
3597	/** @todo Any possible ARM32/ARM64 optimizations here? */
3598	ASMAtomicXchgU32(pu32, u32);
3599	}
3600
3601
3602	/**
3603	* Atomically writes an unsigned 32-bit value, unordered.
3604	*
3605	* @param pu32 Pointer to the 32-bit variable to write.
3606	* @param u32 The 32-bit value to assign to *pu32.
3607	*/
3608	DECLINLINE(void) ASMAtomicUoWriteU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
3609	{
3610	Assert(!((uintptr_t)pu32 & 3));
3611	#if ARCH_BITS >= 32
3612	*pu32 = u32;
3613	#else
3614	ASMAtomicXchgU32(pu32, u32);
3615	#endif
3616	}
3617
3618
3619	/**
3620	* Atomically writes a signed 32-bit value, ordered.
3621	*
3622	* @param pi32 Pointer to the 32-bit variable to write.
3623	* @param i32 The 32-bit value to assign to *pi32.
3624	*/
3625	DECLINLINE(void) ASMAtomicWriteS32(volatile int32_t RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
3626	{
3627	ASMAtomicXchgS32(pi32, i32);
3628	}
3629
3630
3631	/**
3632	* Atomically writes a signed 32-bit value, unordered.
3633	*
3634	* @param pi32 Pointer to the 32-bit variable to write.
3635	* @param i32 The 32-bit value to assign to *pi32.
3636	*/
3637	DECLINLINE(void) ASMAtomicUoWriteS32(volatile int32_t RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
3638	{
3639	Assert(!((uintptr_t)pi32 & 3));
3640	#if ARCH_BITS >= 32
3641	*pi32 = i32;
3642	#else
3643	ASMAtomicXchgS32(pi32, i32);
3644	#endif
3645	}
3646
3647
3648	/**
3649	* Atomically writes an unsigned 64-bit value, ordered.
3650	*
3651	* @param pu64 Pointer to the 64-bit variable to write.
3652	* @param u64 The 64-bit value to assign to *pu64.
3653	*/
3654	DECLINLINE(void) ASMAtomicWriteU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
3655	{
3656	/** @todo Any possible ARM32/ARM64 optimizations here? */
3657	ASMAtomicXchgU64(pu64, u64);
3658	}
3659
3660
3661	/**
3662	* Atomically writes an unsigned 64-bit value, unordered.
3663	*
3664	* @param pu64 Pointer to the 64-bit variable to write.
3665	* @param u64 The 64-bit value to assign to *pu64.
3666	*/
3667	DECLINLINE(void) ASMAtomicUoWriteU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
3668	{
3669	Assert(!((uintptr_t)pu64 & 7));
3670	#if ARCH_BITS == 64
3671	*pu64 = u64;
3672	#else
3673	ASMAtomicXchgU64(pu64, u64);
3674	#endif
3675	}
3676
3677
3678	/**
3679	* Atomically writes a signed 64-bit value, ordered.
3680	*
3681	* @param pi64 Pointer to the 64-bit variable to write.
3682	* @param i64 The 64-bit value to assign to *pi64.
3683	*/
3684	DECLINLINE(void) ASMAtomicWriteS64(volatile int64_t RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
3685	{
3686	/** @todo Any possible ARM32/ARM64 optimizations here? */
3687	ASMAtomicXchgS64(pi64, i64);
3688	}
3689
3690
3691	/**
3692	* Atomically writes a signed 64-bit value, unordered.
3693	*
3694	* @param pi64 Pointer to the 64-bit variable to write.
3695	* @param i64 The 64-bit value to assign to *pi64.
3696	*/
3697	DECLINLINE(void) ASMAtomicUoWriteS64(volatile int64_t RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
3698	{
3699	Assert(!((uintptr_t)pi64 & 7));
3700	#if ARCH_BITS == 64
3701	*pi64 = i64;
3702	#else
3703	ASMAtomicXchgS64(pi64, i64);
3704	#endif
3705	}
3706
3707
3708	/**
3709	* Atomically writes a size_t value, ordered.
3710	*
3711	* @returns nothing.
3712	* @param pcb Pointer to the size_t variable to write.
3713	* @param cb The value to assign to *pcb.
3714	*/
3715	DECLINLINE(void) ASMAtomicWriteZ(volatile size_t RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF
3716	{
3717	#if ARCH_BITS == 64
3718	ASMAtomicWriteU64((uint64_t volatile *)pcb, cb);
3719	#elif ARCH_BITS == 32
3720	ASMAtomicWriteU32((uint32_t volatile *)pcb, cb);
3721	#elif ARCH_BITS == 16
3722	AssertCompileSize(size_t, 2);
3723	ASMAtomicWriteU16((uint16_t volatile *)pcb, cb);
3724	#else
3725	# error "Unsupported ARCH_BITS value"
3726	#endif
3727	}
3728
3729
3730	/**
3731	* Atomically writes a size_t value, unordered.
3732	*
3733	* @returns nothing.
3734	* @param pcb Pointer to the size_t variable to write.
3735	* @param cb The value to assign to *pcb.
3736	*/
3737	DECLINLINE(void) ASMAtomicUoWriteZ(volatile size_t RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF
3738	{
3739	#if ARCH_BITS == 64
3740	ASMAtomicUoWriteU64((uint64_t volatile *)pcb, cb);
3741	#elif ARCH_BITS == 32
3742	ASMAtomicUoWriteU32((uint32_t volatile *)pcb, cb);
3743	#elif ARCH_BITS == 16
3744	AssertCompileSize(size_t, 2);
3745	ASMAtomicUoWriteU16((uint16_t volatile *)pcb, cb);
3746	#else
3747	# error "Unsupported ARCH_BITS value"
3748	#endif
3749	}
3750
3751
3752	/**
3753	* Atomically writes a boolean value, unordered.
3754	*
3755	* @param pf Pointer to the boolean variable to write.
3756	* @param f The boolean value to assign to *pf.
3757	*/
3758	DECLINLINE(void) ASMAtomicWriteBool(volatile bool RT_FAR *pf, bool f) RT_NOTHROW_DEF
3759	{
3760	ASMAtomicWriteU8((uint8_t volatile RT_FAR *)pf, f);
3761	}
3762
3763
3764	/**
3765	* Atomically writes a boolean value, unordered.
3766	*
3767	* @param pf Pointer to the boolean variable to write.
3768	* @param f The boolean value to assign to *pf.
3769	*/
3770	DECLINLINE(void) ASMAtomicUoWriteBool(volatile bool RT_FAR *pf, bool f) RT_NOTHROW_DEF
3771	{
3772	pf = f; / byte writes are atomic on x86 */
3773	}
3774
3775
3776	/**
3777	* Atomically writes a pointer value, ordered.
3778	*
3779	* @param ppv Pointer to the pointer variable to write.
3780	* @param pv The pointer value to assign to *ppv.
3781	*/
3782	DECLINLINE(void) ASMAtomicWritePtrVoid(void RT_FAR * volatile RT_FAR ppv, const void pv) RT_NOTHROW_DEF
3783	{
3784	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
3785	ASMAtomicWriteU32((volatile uint32_t RT_FAR )(void RT_FAR )ppv, (uint32_t)pv);
3786	#elif ARCH_BITS == 64
3787	ASMAtomicWriteU64((volatile uint64_t RT_FAR )(void RT_FAR )ppv, (uint64_t)pv);
3788	#else
3789	# error "ARCH_BITS is bogus"
3790	#endif
3791	}
3792
3793
3794	/**
3795	* Atomically writes a pointer value, unordered.
3796	*
3797	* @param ppv Pointer to the pointer variable to write.
3798	* @param pv The pointer value to assign to *ppv.
3799	*/
3800	DECLINLINE(void) ASMAtomicUoWritePtrVoid(void RT_FAR * volatile RT_FAR ppv, const void pv) RT_NOTHROW_DEF
3801	{
3802	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
3803	ASMAtomicUoWriteU32((volatile uint32_t RT_FAR )(void RT_FAR )ppv, (uint32_t)pv);
3804	#elif ARCH_BITS == 64
3805	ASMAtomicUoWriteU64((volatile uint64_t RT_FAR )(void RT_FAR )ppv, (uint64_t)pv);
3806	#else
3807	# error "ARCH_BITS is bogus"
3808	#endif
3809	}
3810
3811
3812	/**
3813	* Atomically writes a pointer value, ordered.
3814	*
3815	* @param ppv Pointer to the pointer variable to write.
3816	* @param pv The pointer value to assign to *ppv. If NULL use
3817	* ASMAtomicWriteNullPtr or you'll land in trouble.
3818	*
3819	* @remarks This is relatively type safe on GCC platforms when @a pv isn't
3820	* NULL.
3821	*/
3822	#ifdef __GNUC__
3823	# define ASMAtomicWritePtr(ppv, pv) \
3824	do \
3825	{ \
3826	__typeof__((ppv)) volatile RT_FAR const ppvTypeChecked = (ppv); \
3827	__typeof__(*(ppv)) const pvTypeChecked = (pv); \
3828	\
3829	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
3830	AssertCompile(sizeof(pv) == sizeof(void RT_FAR *)); \
3831	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
3832	\
3833	ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR )(ppvTypeChecked), (void RT_FAR )(pvTypeChecked)); \
3834	} while (0)
3835	#else
3836	# define ASMAtomicWritePtr(ppv, pv) \
3837	do \
3838	{ \
3839	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
3840	AssertCompile(sizeof(pv) == sizeof(void RT_FAR *)); \
3841	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
3842	\
3843	ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR )(ppv), (void RT_FAR )(pv)); \
3844	} while (0)
3845	#endif
3846
3847
3848	/**
3849	* Atomically sets a pointer to NULL, ordered.
3850	*
3851	* @param ppv Pointer to the pointer variable that should be set to NULL.
3852	*
3853	* @remarks This is relatively type safe on GCC platforms.
3854	*/
3855	#if RT_GNUC_PREREQ(4, 2)
3856	# define ASMAtomicWriteNullPtr(ppv) \
3857	do \
3858	{ \
3859	__typeof__((ppv)) const ppvTypeChecked = (ppv); \
3860	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
3861	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
3862	ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR *)(ppvTypeChecked), NULL); \
3863	} while (0)
3864	#else
3865	# define ASMAtomicWriteNullPtr(ppv) \
3866	do \
3867	{ \
3868	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
3869	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
3870	ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR *)(ppv), NULL); \
3871	} while (0)
3872	#endif
3873
3874
3875	/**
3876	* Atomically writes a pointer value, unordered.
3877	*
3878	* @returns Current *pv value
3879	* @param ppv Pointer to the pointer variable.
3880	* @param pv The pointer value to assign to *ppv. If NULL use
3881	* ASMAtomicUoWriteNullPtr or you'll land in trouble.
3882	*
3883	* @remarks This is relatively type safe on GCC platforms when @a pv isn't
3884	* NULL.
3885	*/
3886	#if RT_GNUC_PREREQ(4, 2)
3887	# define ASMAtomicUoWritePtr(ppv, pv) \
3888	do \
3889	{ \
3890	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
3891	__typeof__(*(ppv)) const pvTypeChecked = (pv); \
3892	\
3893	AssertCompile(sizeof(ppv) == sizeof(void )); \
3894	AssertCompile(sizeof(pv) == sizeof(void *)); \
3895	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
3896	\
3897	*(ppvTypeChecked) = pvTypeChecked; \
3898	} while (0)
3899	#else
3900	# define ASMAtomicUoWritePtr(ppv, pv) \
3901	do \
3902	{ \
3903	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
3904	AssertCompile(sizeof(pv) == sizeof(void RT_FAR *)); \
3905	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
3906	*(ppv) = pv; \
3907	} while (0)
3908	#endif
3909
3910
3911	/**
3912	* Atomically sets a pointer to NULL, unordered.
3913	*
3914	* @param ppv Pointer to the pointer variable that should be set to NULL.
3915	*
3916	* @remarks This is relatively type safe on GCC platforms.
3917	*/
3918	#ifdef __GNUC__
3919	# define ASMAtomicUoWriteNullPtr(ppv) \
3920	do \
3921	{ \
3922	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
3923	AssertCompile(sizeof(ppv) == sizeof(void )); \
3924	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
3925	*(ppvTypeChecked) = NULL; \
3926	} while (0)
3927	#else
3928	# define ASMAtomicUoWriteNullPtr(ppv) \
3929	do \
3930	{ \
3931	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
3932	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
3933	*(ppv) = NULL; \
3934	} while (0)
3935	#endif
3936
3937
3938	/**
3939	* Atomically write a typical IPRT handle value, ordered.
3940	*
3941	* @param ph Pointer to the variable to update.
3942	* @param hNew The value to assign to *ph.
3943	*
3944	* @remarks This doesn't currently work for all handles (like RTFILE).
3945	*/
3946	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
3947	# define ASMAtomicWriteHandle(ph, hNew) \
3948	do { \
3949	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
3950	ASMAtomicWriteU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)(hNew)); \
3951	} while (0)
3952	#elif HC_ARCH_BITS == 64
3953	# define ASMAtomicWriteHandle(ph, hNew) \
3954	do { \
3955	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
3956	ASMAtomicWriteU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)(hNew)); \
3957	} while (0)
3958	#else
3959	# error HC_ARCH_BITS
3960	#endif
3961
3962
3963	/**
3964	* Atomically write a typical IPRT handle value, unordered.
3965	*
3966	* @param ph Pointer to the variable to update.
3967	* @param hNew The value to assign to *ph.
3968	*
3969	* @remarks This doesn't currently work for all handles (like RTFILE).
3970	*/
3971	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
3972	# define ASMAtomicUoWriteHandle(ph, hNew) \
3973	do { \
3974	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
3975	ASMAtomicUoWriteU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)hNew); \
3976	} while (0)
3977	#elif HC_ARCH_BITS == 64
3978	# define ASMAtomicUoWriteHandle(ph, hNew) \
3979	do { \
3980	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
3981	ASMAtomicUoWriteU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)hNew); \
3982	} while (0)
3983	#else
3984	# error HC_ARCH_BITS
3985	#endif
3986
3987
3988	/**
3989	* Atomically write a value which size might differ
3990	* between platforms or compilers, ordered.
3991	*
3992	* @param pu Pointer to the variable to update.
3993	* @param uNew The value to assign to *pu.
3994	*/
3995	#define ASMAtomicWriteSize(pu, uNew) \
3996	do { \
3997	switch (sizeof(*(pu))) { \
3998	case 1: ASMAtomicWriteU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu), (uint8_t )(uNew)); break; \
3999	case 2: ASMAtomicWriteU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu), (uint16_t)(uNew)); break; \
4000	case 4: ASMAtomicWriteU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
4001	case 8: ASMAtomicWriteU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
4002	default: AssertMsgFailed(("ASMAtomicWriteSize: size %d is not supported\n", sizeof(*(pu)))); \
4003	} \
4004	} while (0)
4005
4006	/**
4007	* Atomically write a value which size might differ
4008	* between platforms or compilers, unordered.
4009	*
4010	* @param pu Pointer to the variable to update.
4011	* @param uNew The value to assign to *pu.
4012	*/
4013	#define ASMAtomicUoWriteSize(pu, uNew) \
4014	do { \
4015	switch (sizeof(*(pu))) { \
4016	case 1: ASMAtomicUoWriteU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu), (uint8_t )(uNew)); break; \
4017	case 2: ASMAtomicUoWriteU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu), (uint16_t)(uNew)); break; \
4018	case 4: ASMAtomicUoWriteU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
4019	case 8: ASMAtomicUoWriteU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
4020	default: AssertMsgFailed(("ASMAtomicWriteSize: size %d is not supported\n", sizeof(*(pu)))); \
4021	} \
4022	} while (0)
4023
4024
4025
4026	/**
4027	* Atomically exchanges and adds to a 16-bit value, ordered.
4028	*
4029	* @returns The old value.
4030	* @param pu16 Pointer to the value.
4031	* @param u16 Number to add.
4032	*
4033	* @remarks Currently not implemented, just to make 16-bit code happy.
4034	* @remarks x86: Requires a 486 or later.
4035	*/
4036	RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMAtomicAddU16(uint16_t volatile RT_FAR *pu16, uint32_t u16) RT_NOTHROW_PROTO;
4037
4038
4039	/**
4040	* Atomically exchanges and adds to a 32-bit value, ordered.
4041	*
4042	* @returns The old value.
4043	* @param pu32 Pointer to the value.
4044	* @param u32 Number to add.
4045	*
4046	* @remarks x86: Requires a 486 or later.
4047	*/
4048	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4049	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicAddU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
4050	#else
4051	DECLINLINE(uint32_t) ASMAtomicAddU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
4052	{
4053	# if RT_INLINE_ASM_USES_INTRIN
4054	u32 = _InterlockedExchangeAdd((long RT_FAR *)pu32, u32);
4055	return u32;
4056
4057	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
4058	# if RT_INLINE_ASM_GNU_STYLE
4059	__asm__ __volatile__("lock; xaddl %0, %1\n\t"
4060	: "=r" (u32)
4061	, "=m" (*pu32)
4062	: "0" (u32)
4063	, "m" (*pu32)
4064	: "memory"
4065	, "cc");
4066	return u32;
4067	# else
4068	__asm
4069	{
4070	mov eax, [u32]
4071	# ifdef RT_ARCH_AMD64
4072	mov rdx, [pu32]
4073	lock xadd [rdx], eax
4074	# else
4075	mov edx, [pu32]
4076	lock xadd [edx], eax
4077	# endif
4078	mov [u32], eax
4079	}
4080	return u32;
4081	# endif
4082
4083	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4084	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicAddU32, pu32, DMB_SY,
4085	"add %w[uNew], %w[uOld], %w[uVal]\n\t",
4086	"add %[uNew], %[uOld], %[uVal]\n\t",
4087	[uVal] "r" (u32));
4088	return u32OldRet;
4089
4090	# else
4091	# error "Port me"
4092	# endif
4093	}
4094	#endif
4095
4096
4097	/**
4098	* Atomically exchanges and adds to a signed 32-bit value, ordered.
4099	*
4100	* @returns The old value.
4101	* @param pi32 Pointer to the value.
4102	* @param i32 Number to add.
4103	*
4104	* @remarks x86: Requires a 486 or later.
4105	*/
4106	DECLINLINE(int32_t) ASMAtomicAddS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
4107	{
4108	return (int32_t)ASMAtomicAddU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
4109	}
4110
4111
4112	/**
4113	* Atomically exchanges and adds to a 64-bit value, ordered.
4114	*
4115	* @returns The old value.
4116	* @param pu64 Pointer to the value.
4117	* @param u64 Number to add.
4118	*
4119	* @remarks x86: Requires a Pentium or later.
4120	*/
4121	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4122	DECLASM(uint64_t) ASMAtomicAddU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
4123	#else
4124	DECLINLINE(uint64_t) ASMAtomicAddU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
4125	{
4126	# if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64)
4127	u64 = _InterlockedExchangeAdd64((__int64 RT_FAR *)pu64, u64);
4128	return u64;
4129
4130	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
4131	__asm__ __volatile__("lock; xaddq %0, %1\n\t"
4132	: "=r" (u64)
4133	, "=m" (*pu64)
4134	: "0" (u64)
4135	, "m" (*pu64)
4136	: "memory"
4137	, "cc");
4138	return u64;
4139
4140	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4141	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_64(ASMAtomicAddU64, pu64, DMB_SY,
4142	"add %[uNew], %[uOld], %[uVal]\n\t"
4143	,
4144	"add %[uNew], %[uOld], %[uVal]\n\t"
4145	"adc %H[uNew], %H[uOld], %H[uVal]\n\t",
4146	[uVal] "r" (u64));
4147	return u64OldRet;
4148
4149	# else
4150	uint64_t u64Old;
4151	for (;;)
4152	{
4153	uint64_t u64New;
4154	u64Old = ASMAtomicUoReadU64(pu64);
4155	u64New = u64Old + u64;
4156	if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old))
4157	break;
4158	ASMNopPause();
4159	}
4160	return u64Old;
4161	# endif
4162	}
4163	#endif
4164
4165
4166	/**
4167	* Atomically exchanges and adds to a signed 64-bit value, ordered.
4168	*
4169	* @returns The old value.
4170	* @param pi64 Pointer to the value.
4171	* @param i64 Number to add.
4172	*
4173	* @remarks x86: Requires a Pentium or later.
4174	*/
4175	DECLINLINE(int64_t) ASMAtomicAddS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
4176	{
4177	return (int64_t)ASMAtomicAddU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64);
4178	}
4179
4180
4181	/**
4182	* Atomically exchanges and adds to a size_t value, ordered.
4183	*
4184	* @returns The old value.
4185	* @param pcb Pointer to the size_t value.
4186	* @param cb Number to add.
4187	*/
4188	DECLINLINE(size_t) ASMAtomicAddZ(size_t volatile RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF
4189	{
4190	#if ARCH_BITS == 64
4191	AssertCompileSize(size_t, 8);
4192	return ASMAtomicAddU64((uint64_t volatile RT_FAR *)pcb, cb);
4193	#elif ARCH_BITS == 32
4194	AssertCompileSize(size_t, 4);
4195	return ASMAtomicAddU32((uint32_t volatile RT_FAR *)pcb, cb);
4196	#elif ARCH_BITS == 16
4197	AssertCompileSize(size_t, 2);
4198	return ASMAtomicAddU16((uint16_t volatile RT_FAR *)pcb, cb);
4199	#else
4200	# error "Unsupported ARCH_BITS value"
4201	#endif
4202	}
4203
4204
4205	/**
4206	* Atomically exchanges and adds a value which size might differ between
4207	* platforms or compilers, ordered.
4208	*
4209	* @param pu Pointer to the variable to update.
4210	* @param uNew The value to add to *pu.
4211	* @param puOld Where to store the old value.
4212	*/
4213	#define ASMAtomicAddSize(pu, uNew, puOld) \
4214	do { \
4215	switch (sizeof(*(pu))) { \
4216	case 4: (uint32_t )(puOld) = ASMAtomicAddU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
4217	case 8: (uint64_t )(puOld) = ASMAtomicAddU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
4218	default: AssertMsgFailed(("ASMAtomicAddSize: size %d is not supported\n", sizeof(*(pu)))); \
4219	} \
4220	} while (0)
4221
4222
4223
4224	/**
4225	* Atomically exchanges and subtracts to an unsigned 16-bit value, ordered.
4226	*
4227	* @returns The old value.
4228	* @param pu16 Pointer to the value.
4229	* @param u16 Number to subtract.
4230	*
4231	* @remarks x86: Requires a 486 or later.
4232	*/
4233	DECLINLINE(uint16_t) ASMAtomicSubU16(uint16_t volatile RT_FAR *pu16, uint32_t u16) RT_NOTHROW_DEF
4234	{
4235	return ASMAtomicAddU16(pu16, (uint16_t)-(int16_t)u16);
4236	}
4237
4238
4239	/**
4240	* Atomically exchanges and subtracts to a signed 16-bit value, ordered.
4241	*
4242	* @returns The old value.
4243	* @param pi16 Pointer to the value.
4244	* @param i16 Number to subtract.
4245	*
4246	* @remarks x86: Requires a 486 or later.
4247	*/
4248	DECLINLINE(int16_t) ASMAtomicSubS16(int16_t volatile RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF
4249	{
4250	return (int16_t)ASMAtomicAddU16((uint16_t volatile RT_FAR *)pi16, (uint16_t)-i16);
4251	}
4252
4253
4254	/**
4255	* Atomically exchanges and subtracts to an unsigned 32-bit value, ordered.
4256	*
4257	* @returns The old value.
4258	* @param pu32 Pointer to the value.
4259	* @param u32 Number to subtract.
4260	*
4261	* @remarks x86: Requires a 486 or later.
4262	*/
4263	DECLINLINE(uint32_t) ASMAtomicSubU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
4264	{
4265	return ASMAtomicAddU32(pu32, (uint32_t)-(int32_t)u32);
4266	}
4267
4268
4269	/**
4270	* Atomically exchanges and subtracts to a signed 32-bit value, ordered.
4271	*
4272	* @returns The old value.
4273	* @param pi32 Pointer to the value.
4274	* @param i32 Number to subtract.
4275	*
4276	* @remarks x86: Requires a 486 or later.
4277	*/
4278	DECLINLINE(int32_t) ASMAtomicSubS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
4279	{
4280	return (int32_t)ASMAtomicAddU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)-i32);
4281	}
4282
4283
4284	/**
4285	* Atomically exchanges and subtracts to an unsigned 64-bit value, ordered.
4286	*
4287	* @returns The old value.
4288	* @param pu64 Pointer to the value.
4289	* @param u64 Number to subtract.
4290	*
4291	* @remarks x86: Requires a Pentium or later.
4292	*/
4293	DECLINLINE(uint64_t) ASMAtomicSubU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
4294	{
4295	return ASMAtomicAddU64(pu64, (uint64_t)-(int64_t)u64);
4296	}
4297
4298
4299	/**
4300	* Atomically exchanges and subtracts to a signed 64-bit value, ordered.
4301	*
4302	* @returns The old value.
4303	* @param pi64 Pointer to the value.
4304	* @param i64 Number to subtract.
4305	*
4306	* @remarks x86: Requires a Pentium or later.
4307	*/
4308	DECLINLINE(int64_t) ASMAtomicSubS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
4309	{
4310	return (int64_t)ASMAtomicAddU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)-i64);
4311	}
4312
4313
4314	/**
4315	* Atomically exchanges and subtracts to a size_t value, ordered.
4316	*
4317	* @returns The old value.
4318	* @param pcb Pointer to the size_t value.
4319	* @param cb Number to subtract.
4320	*
4321	* @remarks x86: Requires a 486 or later.
4322	*/
4323	DECLINLINE(size_t) ASMAtomicSubZ(size_t volatile RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF
4324	{
4325	#if ARCH_BITS == 64
4326	return ASMAtomicSubU64((uint64_t volatile RT_FAR *)pcb, cb);
4327	#elif ARCH_BITS == 32
4328	return ASMAtomicSubU32((uint32_t volatile RT_FAR *)pcb, cb);
4329	#elif ARCH_BITS == 16
4330	AssertCompileSize(size_t, 2);
4331	return ASMAtomicSubU16((uint16_t volatile RT_FAR *)pcb, cb);
4332	#else
4333	# error "Unsupported ARCH_BITS value"
4334	#endif
4335	}
4336
4337
4338	/**
4339	* Atomically exchanges and subtracts a value which size might differ between
4340	* platforms or compilers, ordered.
4341	*
4342	* @param pu Pointer to the variable to update.
4343	* @param uNew The value to subtract to *pu.
4344	* @param puOld Where to store the old value.
4345	*
4346	* @remarks x86: Requires a 486 or later.
4347	*/
4348	#define ASMAtomicSubSize(pu, uNew, puOld) \
4349	do { \
4350	switch (sizeof(*(pu))) { \
4351	case 4: (uint32_t RT_FAR )(puOld) = ASMAtomicSubU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
4352	case 8: (uint64_t RT_FAR )(puOld) = ASMAtomicSubU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
4353	default: AssertMsgFailed(("ASMAtomicSubSize: size %d is not supported\n", sizeof(*(pu)))); \
4354	} \
4355	} while (0)
4356
4357
4358
4359	/**
4360	* Atomically increment a 16-bit value, ordered.
4361	*
4362	* @returns The new value.
4363	* @param pu16 Pointer to the value to increment.
4364	* @remarks Not implemented. Just to make 16-bit code happy.
4365	*
4366	* @remarks x86: Requires a 486 or later.
4367	*/
4368	RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMAtomicIncU16(uint16_t volatile RT_FAR *pu16) RT_NOTHROW_PROTO;
4369
4370
4371	/**
4372	* Atomically increment a 32-bit value, ordered.
4373	*
4374	* @returns The new value.
4375	* @param pu32 Pointer to the value to increment.
4376	*
4377	* @remarks x86: Requires a 486 or later.
4378	*/
4379	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4380	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO;
4381	#else
4382	DECLINLINE(uint32_t) ASMAtomicIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF
4383	{
4384	# if RT_INLINE_ASM_USES_INTRIN
4385	return (uint32_t)_InterlockedIncrement((long RT_FAR *)pu32);
4386
4387	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
4388	# if RT_INLINE_ASM_GNU_STYLE
4389	uint32_t u32;
4390	__asm__ __volatile__("lock; xaddl %0, %1\n\t"
4391	: "=r" (u32)
4392	, "=m" (*pu32)
4393	: "0" (1)
4394	, "m" (*pu32)
4395	: "memory"
4396	, "cc");
4397	return u32+1;
4398	# else
4399	__asm
4400	{
4401	mov eax, 1
4402	# ifdef RT_ARCH_AMD64
4403	mov rdx, [pu32]
4404	lock xadd [rdx], eax
4405	# else
4406	mov edx, [pu32]
4407	lock xadd [edx], eax
4408	# endif
4409	mov u32, eax
4410	}
4411	return u32+1;
4412	# endif
4413
4414	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4415	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicIncU32, pu32, DMB_SY,
4416	"add %w[uNew], %w[uNew], #1\n\t",
4417	"add %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */,
4418	"X" (0) /* dummy */);
4419	return u32NewRet;
4420
4421	# else
4422	return ASMAtomicAddU32(pu32, 1) + 1;
4423	# endif
4424	}
4425	#endif
4426
4427
4428	/**
4429	* Atomically increment a signed 32-bit value, ordered.
4430	*
4431	* @returns The new value.
4432	* @param pi32 Pointer to the value to increment.
4433	*
4434	* @remarks x86: Requires a 486 or later.
4435	*/
4436	DECLINLINE(int32_t) ASMAtomicIncS32(int32_t volatile RT_FAR *pi32) RT_NOTHROW_DEF
4437	{
4438	return (int32_t)ASMAtomicIncU32((uint32_t volatile RT_FAR *)pi32);
4439	}
4440
4441
4442	/**
4443	* Atomically increment a 64-bit value, ordered.
4444	*
4445	* @returns The new value.
4446	* @param pu64 Pointer to the value to increment.
4447	*
4448	* @remarks x86: Requires a Pentium or later.
4449	*/
4450	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4451	DECLASM(uint64_t) ASMAtomicIncU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_PROTO;
4452	#else
4453	DECLINLINE(uint64_t) ASMAtomicIncU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_DEF
4454	{
4455	# if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64)
4456	return (uint64_t)_InterlockedIncrement64((__int64 RT_FAR *)pu64);
4457
4458	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
4459	uint64_t u64;
4460	__asm__ __volatile__("lock; xaddq %0, %1\n\t"
4461	: "=r" (u64)
4462	, "=m" (*pu64)
4463	: "0" (1)
4464	, "m" (*pu64)
4465	: "memory"
4466	, "cc");
4467	return u64 + 1;
4468
4469	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4470	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicIncU64, pu64, DMB_SY,
4471	"add %[uNew], %[uNew], #1\n\t"
4472	,
4473	"add %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */
4474	"adc %H[uNew], %H[uNew], %[uZeroVal]\n\t",
4475	RTASM_ARM_PICK_6432("X" (0) /* dummy */, [uZeroVal] "r" (0)) );
4476	return u64NewRet;
4477
4478	# else
4479	return ASMAtomicAddU64(pu64, 1) + 1;
4480	# endif
4481	}
4482	#endif
4483
4484
4485	/**
4486	* Atomically increment a signed 64-bit value, ordered.
4487	*
4488	* @returns The new value.
4489	* @param pi64 Pointer to the value to increment.
4490	*
4491	* @remarks x86: Requires a Pentium or later.
4492	*/
4493	DECLINLINE(int64_t) ASMAtomicIncS64(int64_t volatile RT_FAR *pi64) RT_NOTHROW_DEF
4494	{
4495	return (int64_t)ASMAtomicIncU64((uint64_t volatile RT_FAR *)pi64);
4496	}
4497
4498
4499	/**
4500	* Atomically increment a size_t value, ordered.
4501	*
4502	* @returns The new value.
4503	* @param pcb Pointer to the value to increment.
4504	*
4505	* @remarks x86: Requires a 486 or later.
4506	*/
4507	DECLINLINE(size_t) ASMAtomicIncZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF
4508	{
4509	#if ARCH_BITS == 64
4510	return ASMAtomicIncU64((uint64_t volatile RT_FAR *)pcb);
4511	#elif ARCH_BITS == 32
4512	return ASMAtomicIncU32((uint32_t volatile RT_FAR *)pcb);
4513	#elif ARCH_BITS == 16
4514	return ASMAtomicIncU16((uint16_t volatile RT_FAR *)pcb);
4515	#else
4516	# error "Unsupported ARCH_BITS value"
4517	#endif
4518	}
4519
4520
4521
4522	/**
4523	* Atomically decrement an unsigned 32-bit value, ordered.
4524	*
4525	* @returns The new value.
4526	* @param pu16 Pointer to the value to decrement.
4527	* @remarks Not implemented. Just to make 16-bit code happy.
4528	*
4529	* @remarks x86: Requires a 486 or later.
4530	*/
4531	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicDecU16(uint16_t volatile RT_FAR *pu16) RT_NOTHROW_PROTO;
4532
4533
4534	/**
4535	* Atomically decrement an unsigned 32-bit value, ordered.
4536	*
4537	* @returns The new value.
4538	* @param pu32 Pointer to the value to decrement.
4539	*
4540	* @remarks x86: Requires a 486 or later.
4541	*/
4542	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4543	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO;
4544	#else
4545	DECLINLINE(uint32_t) ASMAtomicDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF
4546	{
4547	# if RT_INLINE_ASM_USES_INTRIN
4548	return (uint32_t)_InterlockedDecrement((long RT_FAR *)pu32);
4549
4550	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
4551	# if RT_INLINE_ASM_GNU_STYLE
4552	uint32_t u32;
4553	__asm__ __volatile__("lock; xaddl %0, %1\n\t"
4554	: "=r" (u32)
4555	, "=m" (*pu32)
4556	: "0" (-1)
4557	, "m" (*pu32)
4558	: "memory"
4559	, "cc");
4560	return u32-1;
4561	# else
4562	uint32_t u32;
4563	__asm
4564	{
4565	mov eax, -1
4566	# ifdef RT_ARCH_AMD64
4567	mov rdx, [pu32]
4568	lock xadd [rdx], eax
4569	# else
4570	mov edx, [pu32]
4571	lock xadd [edx], eax
4572	# endif
4573	mov u32, eax
4574	}
4575	return u32-1;
4576	# endif
4577
4578	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4579	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicDecU32, pu32, DMB_SY,
4580	"sub %w[uNew], %w[uNew], #1\n\t",
4581	"sub %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */,
4582	"X" (0) /* dummy */);
4583	return u32NewRet;
4584
4585	# else
4586	return ASMAtomicSubU32(pu32, 1) - (uint32_t)1;
4587	# endif
4588	}
4589	#endif
4590
4591
4592	/**
4593	* Atomically decrement a signed 32-bit value, ordered.
4594	*
4595	* @returns The new value.
4596	* @param pi32 Pointer to the value to decrement.
4597	*
4598	* @remarks x86: Requires a 486 or later.
4599	*/
4600	DECLINLINE(int32_t) ASMAtomicDecS32(int32_t volatile RT_FAR *pi32) RT_NOTHROW_DEF
4601	{
4602	return (int32_t)ASMAtomicDecU32((uint32_t volatile RT_FAR *)pi32);
4603	}
4604
4605
4606	/**
4607	* Atomically decrement an unsigned 64-bit value, ordered.
4608	*
4609	* @returns The new value.
4610	* @param pu64 Pointer to the value to decrement.
4611	*
4612	* @remarks x86: Requires a Pentium or later.
4613	*/
4614	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4615	RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicDecU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_PROTO;
4616	#else
4617	DECLINLINE(uint64_t) ASMAtomicDecU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_DEF
4618	{
4619	# if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64)
4620	return (uint64_t)_InterlockedDecrement64((__int64 volatile RT_FAR *)pu64);
4621
4622	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
4623	uint64_t u64;
4624	__asm__ __volatile__("lock; xaddq %q0, %1\n\t"
4625	: "=r" (u64)
4626	, "=m" (*pu64)
4627	: "0" (~(uint64_t)0)
4628	, "m" (*pu64)
4629	: "memory"
4630	, "cc");
4631	return u64-1;
4632
4633	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4634	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicDecU64, pu64, DMB_SY,
4635	"sub %[uNew], %[uNew], #1\n\t"
4636	,
4637	"sub %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */
4638	"sbc %H[uNew], %H[uNew], %[uZeroVal]\n\t",
4639	RTASM_ARM_PICK_6432("X" (0) /* dummy */, [uZeroVal] "r" (0)) );
4640	return u64NewRet;
4641
4642	# else
4643	return ASMAtomicAddU64(pu64, UINT64_MAX) - 1;
4644	# endif
4645	}
4646	#endif
4647
4648
4649	/**
4650	* Atomically decrement a signed 64-bit value, ordered.
4651	*
4652	* @returns The new value.
4653	* @param pi64 Pointer to the value to decrement.
4654	*
4655	* @remarks x86: Requires a Pentium or later.
4656	*/
4657	DECLINLINE(int64_t) ASMAtomicDecS64(int64_t volatile RT_FAR *pi64) RT_NOTHROW_DEF
4658	{
4659	return (int64_t)ASMAtomicDecU64((uint64_t volatile RT_FAR *)pi64);
4660	}
4661
4662
4663	/**
4664	* Atomically decrement a size_t value, ordered.
4665	*
4666	* @returns The new value.
4667	* @param pcb Pointer to the value to decrement.
4668	*
4669	* @remarks x86: Requires a 486 or later.
4670	*/
4671	DECLINLINE(size_t) ASMAtomicDecZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF
4672	{
4673	#if ARCH_BITS == 64
4674	return ASMAtomicDecU64((uint64_t volatile RT_FAR *)pcb);
4675	#elif ARCH_BITS == 32
4676	return ASMAtomicDecU32((uint32_t volatile RT_FAR *)pcb);
4677	#elif ARCH_BITS == 16
4678	return ASMAtomicDecU16((uint16_t volatile RT_FAR *)pcb);
4679	#else
4680	# error "Unsupported ARCH_BITS value"
4681	#endif
4682	}
4683
4684
4685	/**
4686	* Atomically Or an unsigned 32-bit value, ordered.
4687	*
4688	* @param pu32 Pointer to the pointer variable to OR u32 with.
4689	* @param u32 The value to OR *pu32 with.
4690	*
4691	* @remarks x86: Requires a 386 or later.
4692	*/
4693	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4694	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
4695	#else
4696	DECLINLINE(void) ASMAtomicOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
4697	{
4698	# if RT_INLINE_ASM_USES_INTRIN
4699	_InterlockedOr((long volatile RT_FAR *)pu32, (long)u32);
4700
4701	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
4702	# if RT_INLINE_ASM_GNU_STYLE
4703	__asm__ __volatile__("lock; orl %1, %0\n\t"
4704	: "=m" (*pu32)
4705	: "ir" (u32)
4706	, "m" (*pu32)
4707	: "cc");
4708	# else
4709	__asm
4710	{
4711	mov eax, [u32]
4712	# ifdef RT_ARCH_AMD64
4713	mov rdx, [pu32]
4714	lock or [rdx], eax
4715	# else
4716	mov edx, [pu32]
4717	lock or [edx], eax
4718	# endif
4719	}
4720	# endif
4721
4722	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4723	/* For more on Orr see https://en.wikipedia.org/wiki/Orr_(Catch-22) ;-) */
4724	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicOr32, pu32, DMB_SY,
4725	"orr %w[uNew], %w[uNew], %w[uVal]\n\t",
4726	"orr %[uNew], %[uNew], %[uVal]\n\t",
4727	[uVal] "r" (u32));
4728
4729	# else
4730	# error "Port me"
4731	# endif
4732	}
4733	#endif
4734
4735
4736	/**
4737	* Atomically OR an unsigned 32-bit value, ordered, extended version (for bitmap
4738	* fallback).
4739	*
4740	* @returns Old value.
4741	* @param pu32 Pointer to the variable to OR @a u32 with.
4742	* @param u32 The value to OR @a *pu32 with.
4743	*/
4744	DECLINLINE(uint32_t) ASMAtomicOrExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
4745	{
4746	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4747	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicOrEx32, pu32, DMB_SY,
4748	"orr %w[uNew], %w[uOld], %w[uVal]\n\t",
4749	"orr %[uNew], %[uOld], %[uVal]\n\t",
4750	[uVal] "r" (u32));
4751	return u32OldRet;
4752
4753	#else
4754	uint32_t u32RetOld = ASMAtomicUoReadU32(pu32);
4755	uint32_t u32New;
4756	do
4757	u32New = u32RetOld \| u32;
4758	while (!ASMAtomicCmpXchgExU32(pu32, u32New, u32RetOld, &u32RetOld));
4759	return u32RetOld;
4760	#endif
4761	}
4762
4763
4764	/**
4765	* Atomically Or a signed 32-bit value, ordered.
4766	*
4767	* @param pi32 Pointer to the pointer variable to OR u32 with.
4768	* @param i32 The value to OR *pu32 with.
4769	*
4770	* @remarks x86: Requires a 386 or later.
4771	*/
4772	DECLINLINE(void) ASMAtomicOrS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
4773	{
4774	ASMAtomicOrU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
4775	}
4776
4777
4778	/**
4779	* Atomically Or an unsigned 64-bit value, ordered.
4780	*
4781	* @param pu64 Pointer to the pointer variable to OR u64 with.
4782	* @param u64 The value to OR *pu64 with.
4783	*
4784	* @remarks x86: Requires a Pentium or later.
4785	*/
4786	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4787	DECLASM(void) ASMAtomicOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
4788	#else
4789	DECLINLINE(void) ASMAtomicOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
4790	{
4791	# if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64)
4792	_InterlockedOr64((__int64 volatile RT_FAR *)pu64, (__int64)u64);
4793
4794	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
4795	__asm__ __volatile__("lock; orq %1, %q0\n\t"
4796	: "=m" (*pu64)
4797	: "r" (u64)
4798	, "m" (*pu64)
4799	: "cc");
4800
4801	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4802	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicOrU64, pu64, DMB_SY,
4803	"orr %[uNew], %[uNew], %[uVal]\n\t"
4804	,
4805	"orr %[uNew], %[uNew], %[uVal]\n\t"
4806	"orr %H[uNew], %H[uNew], %H[uVal]\n\t",
4807	[uVal] "r" (u64));
4808
4809	# else
4810	for (;;)
4811	{
4812	uint64_t u64Old = ASMAtomicUoReadU64(pu64);
4813	uint64_t u64New = u64Old \| u64;
4814	if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old))
4815	break;
4816	ASMNopPause();
4817	}
4818	# endif
4819	}
4820	#endif
4821
4822
4823	/**
4824	* Atomically Or a signed 64-bit value, ordered.
4825	*
4826	* @param pi64 Pointer to the pointer variable to OR u64 with.
4827	* @param i64 The value to OR *pu64 with.
4828	*
4829	* @remarks x86: Requires a Pentium or later.
4830	*/
4831	DECLINLINE(void) ASMAtomicOrS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
4832	{
4833	ASMAtomicOrU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64);
4834	}
4835
4836
4837	/**
4838	* Atomically And an unsigned 32-bit value, ordered.
4839	*
4840	* @param pu32 Pointer to the pointer variable to AND u32 with.
4841	* @param u32 The value to AND *pu32 with.
4842	*
4843	* @remarks x86: Requires a 386 or later.
4844	*/
4845	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4846	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
4847	#else
4848	DECLINLINE(void) ASMAtomicAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
4849	{
4850	# if RT_INLINE_ASM_USES_INTRIN
4851	_InterlockedAnd((long volatile RT_FAR *)pu32, u32);
4852
4853	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
4854	# if RT_INLINE_ASM_GNU_STYLE
4855	__asm__ __volatile__("lock; andl %1, %0\n\t"
4856	: "=m" (*pu32)
4857	: "ir" (u32)
4858	, "m" (*pu32)
4859	: "cc");
4860	# else
4861	__asm
4862	{
4863	mov eax, [u32]
4864	# ifdef RT_ARCH_AMD64
4865	mov rdx, [pu32]
4866	lock and [rdx], eax
4867	# else
4868	mov edx, [pu32]
4869	lock and [edx], eax
4870	# endif
4871	}
4872	# endif
4873
4874	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4875	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicAnd32, pu32, DMB_SY,
4876	"and %w[uNew], %w[uNew], %w[uVal]\n\t",
4877	"and %[uNew], %[uNew], %[uVal]\n\t",
4878	[uVal] "r" (u32));
4879
4880	# else
4881	# error "Port me"
4882	# endif
4883	}
4884	#endif
4885
4886
4887	/**
4888	* Atomically AND an unsigned 32-bit value, ordered, extended version.
4889	*
4890	* @returns Old value.
4891	* @param pu32 Pointer to the variable to AND @a u32 with.
4892	* @param u32 The value to AND @a *pu32 with.
4893	*/
4894	DECLINLINE(uint32_t) ASMAtomicAndExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
4895	{
4896	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4897	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicAndEx32, pu32, DMB_SY,
4898	"and %w[uNew], %w[uOld], %w[uVal]\n\t",
4899	"and %[uNew], %[uOld], %[uVal]\n\t",
4900	[uVal] "r" (u32));
4901	return u32OldRet;
4902
4903	#else
4904	uint32_t u32RetOld = ASMAtomicUoReadU32(pu32);
4905	uint32_t u32New;
4906	do
4907	u32New = u32RetOld & u32;
4908	while (!ASMAtomicCmpXchgExU32(pu32, u32New, u32RetOld, &u32RetOld));
4909	return u32RetOld;
4910	#endif
4911	}
4912
4913
4914	/**
4915	* Atomically And a signed 32-bit value, ordered.
4916	*
4917	* @param pi32 Pointer to the pointer variable to AND i32 with.
4918	* @param i32 The value to AND *pi32 with.
4919	*
4920	* @remarks x86: Requires a 386 or later.
4921	*/
4922	DECLINLINE(void) ASMAtomicAndS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
4923	{
4924	ASMAtomicAndU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
4925	}
4926
4927
4928	/**
4929	* Atomically And an unsigned 64-bit value, ordered.
4930	*
4931	* @param pu64 Pointer to the pointer variable to AND u64 with.
4932	* @param u64 The value to AND *pu64 with.
4933	*
4934	* @remarks x86: Requires a Pentium or later.
4935	*/
4936	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4937	DECLASM(void) ASMAtomicAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
4938	#else
4939	DECLINLINE(void) ASMAtomicAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
4940	{
4941	# if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64)
4942	_InterlockedAnd64((__int64 volatile RT_FAR *)pu64, u64);
4943
4944	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
4945	__asm__ __volatile__("lock; andq %1, %0\n\t"
4946	: "=m" (*pu64)
4947	: "r" (u64)
4948	, "m" (*pu64)
4949	: "cc");
4950
4951	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4952	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicAndU64, pu64, DMB_SY,
4953	"and %[uNew], %[uNew], %[uVal]\n\t"
4954	,
4955	"and %[uNew], %[uNew], %[uVal]\n\t"
4956	"and %H[uNew], %H[uNew], %H[uVal]\n\t",
4957	[uVal] "r" (u64));
4958
4959	# else
4960	for (;;)
4961	{
4962	uint64_t u64Old = ASMAtomicUoReadU64(pu64);
4963	uint64_t u64New = u64Old & u64;
4964	if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old))
4965	break;
4966	ASMNopPause();
4967	}
4968	# endif
4969	}
4970	#endif
4971
4972
4973	/**
4974	* Atomically And a signed 64-bit value, ordered.
4975	*
4976	* @param pi64 Pointer to the pointer variable to AND i64 with.
4977	* @param i64 The value to AND *pi64 with.
4978	*
4979	* @remarks x86: Requires a Pentium or later.
4980	*/
4981	DECLINLINE(void) ASMAtomicAndS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
4982	{
4983	ASMAtomicAndU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64);
4984	}
4985
4986
4987	/**
4988	* Atomically XOR an unsigned 32-bit value and a memory location, ordered.
4989	*
4990	* @param pu32 Pointer to the variable to XOR @a u32 with.
4991	* @param u32 The value to XOR @a *pu32 with.
4992	*
4993	* @remarks x86: Requires a 386 or later.
4994	*/
4995	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4996	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
4997	#else
4998	DECLINLINE(void) ASMAtomicXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
4999	{
5000	# if RT_INLINE_ASM_USES_INTRIN
5001	_InterlockedXor((long volatile RT_FAR *)pu32, u32);
5002
5003	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5004	# if RT_INLINE_ASM_GNU_STYLE
5005	__asm__ __volatile__("lock; xorl %1, %0\n\t"
5006	: "=m" (*pu32)
5007	: "ir" (u32)
5008	, "m" (*pu32)
5009	: "cc");
5010	# else
5011	__asm
5012	{
5013	mov eax, [u32]
5014	# ifdef RT_ARCH_AMD64
5015	mov rdx, [pu32]
5016	lock xor [rdx], eax
5017	# else
5018	mov edx, [pu32]
5019	lock xor [edx], eax
5020	# endif
5021	}
5022	# endif
5023
5024	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5025	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicXor32, pu32, DMB_SY,
5026	"eor %w[uNew], %w[uNew], %w[uVal]\n\t",
5027	"eor %[uNew], %[uNew], %[uVal]\n\t",
5028	[uVal] "r" (u32));
5029
5030	# else
5031	# error "Port me"
5032	# endif
5033	}
5034	#endif
5035
5036
5037	/**
5038	* Atomically XOR an unsigned 32-bit value and a memory location, ordered,
5039	* extended version (for bitmaps).
5040	*
5041	* @returns Old value.
5042	* @param pu32 Pointer to the variable to XOR @a u32 with.
5043	* @param u32 The value to XOR @a *pu32 with.
5044	*/
5045	DECLINLINE(uint32_t) ASMAtomicXorExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5046	{
5047	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5048	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicXorEx32, pu32, DMB_SY,
5049	"eor %w[uNew], %w[uOld], %w[uVal]\n\t",
5050	"eor %[uNew], %[uOld], %[uVal]\n\t",
5051	[uVal] "r" (u32));
5052	return u32OldRet;
5053
5054	#else
5055	uint32_t u32RetOld = ASMAtomicUoReadU32(pu32);
5056	uint32_t u32New;
5057	do
5058	u32New = u32RetOld ^ u32;
5059	while (!ASMAtomicCmpXchgExU32(pu32, u32New, u32RetOld, &u32RetOld));
5060	return u32RetOld;
5061	#endif
5062	}
5063
5064
5065	/**
5066	* Atomically XOR a signed 32-bit value, ordered.
5067	*
5068	* @param pi32 Pointer to the variable to XOR i32 with.
5069	* @param i32 The value to XOR *pi32 with.
5070	*
5071	* @remarks x86: Requires a 386 or later.
5072	*/
5073	DECLINLINE(void) ASMAtomicXorS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
5074	{
5075	ASMAtomicXorU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
5076	}
5077
5078
5079	/**
5080	* Atomically OR an unsigned 32-bit value, unordered but interrupt safe.
5081	*
5082	* @param pu32 Pointer to the pointer variable to OR u32 with.
5083	* @param u32 The value to OR *pu32 with.
5084	*
5085	* @remarks x86: Requires a 386 or later.
5086	*/
5087	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
5088	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicUoOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
5089	#else
5090	DECLINLINE(void) ASMAtomicUoOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5091	{
5092	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5093	# if RT_INLINE_ASM_GNU_STYLE
5094	__asm__ __volatile__("orl %1, %0\n\t"
5095	: "=m" (*pu32)
5096	: "ir" (u32)
5097	, "m" (*pu32)
5098	: "cc");
5099	# else
5100	__asm
5101	{
5102	mov eax, [u32]
5103	# ifdef RT_ARCH_AMD64
5104	mov rdx, [pu32]
5105	or [rdx], eax
5106	# else
5107	mov edx, [pu32]
5108	or [edx], eax
5109	# endif
5110	}
5111	# endif
5112
5113	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5114	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoOrU32, pu32, NO_BARRIER,
5115	"orr %w[uNew], %w[uNew], %w[uVal]\n\t",
5116	"orr %[uNew], %[uNew], %[uVal]\n\t",
5117	[uVal] "r" (u32));
5118
5119	# else
5120	# error "Port me"
5121	# endif
5122	}
5123	#endif
5124
5125
5126	/**
5127	* Atomically OR an unsigned 32-bit value, unordered but interrupt safe,
5128	* extended version (for bitmap fallback).
5129	*
5130	* @returns Old value.
5131	* @param pu32 Pointer to the variable to OR @a u32 with.
5132	* @param u32 The value to OR @a *pu32 with.
5133	*/
5134	DECLINLINE(uint32_t) ASMAtomicUoOrExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5135	{
5136	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5137	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicUoOrExU32, pu32, NO_BARRIER,
5138	"orr %w[uNew], %w[uOld], %w[uVal]\n\t",
5139	"orr %[uNew], %[uOld], %[uVal]\n\t",
5140	[uVal] "r" (u32));
5141	return u32OldRet;
5142
5143	#else
5144	return ASMAtomicOrExU32(pu32, u32); /* (we have no unordered cmpxchg primitive atm.) */
5145	#endif
5146	}
5147
5148
5149	/**
5150	* Atomically OR a signed 32-bit value, unordered.
5151	*
5152	* @param pi32 Pointer to the pointer variable to OR u32 with.
5153	* @param i32 The value to OR *pu32 with.
5154	*
5155	* @remarks x86: Requires a 386 or later.
5156	*/
5157	DECLINLINE(void) ASMAtomicUoOrS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
5158	{
5159	ASMAtomicUoOrU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
5160	}
5161
5162
5163	/**
5164	* Atomically OR an unsigned 64-bit value, unordered.
5165	*
5166	* @param pu64 Pointer to the pointer variable to OR u64 with.
5167	* @param u64 The value to OR *pu64 with.
5168	*
5169	* @remarks x86: Requires a Pentium or later.
5170	*/
5171	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
5172	DECLASM(void) ASMAtomicUoOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
5173	#else
5174	DECLINLINE(void) ASMAtomicUoOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
5175	{
5176	# if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
5177	__asm__ __volatile__("orq %1, %q0\n\t"
5178	: "=m" (*pu64)
5179	: "r" (u64)
5180	, "m" (*pu64)
5181	: "cc");
5182
5183	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5184	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicUoOrU64, pu64, NO_BARRIER,
5185	"orr %[uNew], %[uNew], %[uVal]\n\t"
5186	,
5187	"orr %[uNew], %[uNew], %[uVal]\n\t"
5188	"orr %H[uNew], %H[uNew], %H[uVal]\n\t",
5189	[uVal] "r" (u64));
5190
5191	# else
5192	for (;;)
5193	{
5194	uint64_t u64Old = ASMAtomicUoReadU64(pu64);
5195	uint64_t u64New = u64Old \| u64;
5196	if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old))
5197	break;
5198	ASMNopPause();
5199	}
5200	# endif
5201	}
5202	#endif
5203
5204
5205	/**
5206	* Atomically Or a signed 64-bit value, unordered.
5207	*
5208	* @param pi64 Pointer to the pointer variable to OR u64 with.
5209	* @param i64 The value to OR *pu64 with.
5210	*
5211	* @remarks x86: Requires a Pentium or later.
5212	*/
5213	DECLINLINE(void) ASMAtomicUoOrS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
5214	{
5215	ASMAtomicUoOrU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64);
5216	}
5217
5218
5219	/**
5220	* Atomically And an unsigned 32-bit value, unordered.
5221	*
5222	* @param pu32 Pointer to the pointer variable to AND u32 with.
5223	* @param u32 The value to AND *pu32 with.
5224	*
5225	* @remarks x86: Requires a 386 or later.
5226	*/
5227	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
5228	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicUoAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
5229	#else
5230	DECLINLINE(void) ASMAtomicUoAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5231	{
5232	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5233	# if RT_INLINE_ASM_GNU_STYLE
5234	__asm__ __volatile__("andl %1, %0\n\t"
5235	: "=m" (*pu32)
5236	: "ir" (u32)
5237	, "m" (*pu32)
5238	: "cc");
5239	# else
5240	__asm
5241	{
5242	mov eax, [u32]
5243	# ifdef RT_ARCH_AMD64
5244	mov rdx, [pu32]
5245	and [rdx], eax
5246	# else
5247	mov edx, [pu32]
5248	and [edx], eax
5249	# endif
5250	}
5251	# endif
5252
5253	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5254	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoAnd32, pu32, NO_BARRIER,
5255	"and %w[uNew], %w[uNew], %w[uVal]\n\t",
5256	"and %[uNew], %[uNew], %[uVal]\n\t",
5257	[uVal] "r" (u32));
5258
5259	# else
5260	# error "Port me"
5261	# endif
5262	}
5263	#endif
5264
5265
5266	/**
5267	* Atomically AND an unsigned 32-bit value, unordered, extended version (for
5268	* bitmap fallback).
5269	*
5270	* @returns Old value.
5271	* @param pu32 Pointer to the pointer to AND @a u32 with.
5272	* @param u32 The value to AND @a *pu32 with.
5273	*/
5274	DECLINLINE(uint32_t) ASMAtomicUoAndExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5275	{
5276	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5277	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicUoAndEx32, pu32, NO_BARRIER,
5278	"and %w[uNew], %w[uOld], %w[uVal]\n\t",
5279	"and %[uNew], %[uOld], %[uVal]\n\t",
5280	[uVal] "r" (u32));
5281	return u32OldRet;
5282
5283	#else
5284	return ASMAtomicAndExU32(pu32, u32); /* (we have no unordered cmpxchg primitive atm.) */
5285	#endif
5286	}
5287
5288
5289	/**
5290	* Atomically And a signed 32-bit value, unordered.
5291	*
5292	* @param pi32 Pointer to the pointer variable to AND i32 with.
5293	* @param i32 The value to AND *pi32 with.
5294	*
5295	* @remarks x86: Requires a 386 or later.
5296	*/
5297	DECLINLINE(void) ASMAtomicUoAndS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
5298	{
5299	ASMAtomicUoAndU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
5300	}
5301
5302
5303	/**
5304	* Atomically And an unsigned 64-bit value, unordered.
5305	*
5306	* @param pu64 Pointer to the pointer variable to AND u64 with.
5307	* @param u64 The value to AND *pu64 with.
5308	*
5309	* @remarks x86: Requires a Pentium or later.
5310	*/
5311	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
5312	DECLASM(void) ASMAtomicUoAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
5313	#else
5314	DECLINLINE(void) ASMAtomicUoAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
5315	{
5316	# if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
5317	__asm__ __volatile__("andq %1, %0\n\t"
5318	: "=m" (*pu64)
5319	: "r" (u64)
5320	, "m" (*pu64)
5321	: "cc");
5322
5323	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5324	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicUoAndU64, pu64, NO_BARRIER,
5325	"and %[uNew], %[uNew], %[uVal]\n\t"
5326	,
5327	"and %[uNew], %[uNew], %[uVal]\n\t"
5328	"and %H[uNew], %H[uNew], %H[uVal]\n\t",
5329	[uVal] "r" (u64));
5330
5331	# else
5332	for (;;)
5333	{
5334	uint64_t u64Old = ASMAtomicUoReadU64(pu64);
5335	uint64_t u64New = u64Old & u64;
5336	if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old))
5337	break;
5338	ASMNopPause();
5339	}
5340	# endif
5341	}
5342	#endif
5343
5344
5345	/**
5346	* Atomically And a signed 64-bit value, unordered.
5347	*
5348	* @param pi64 Pointer to the pointer variable to AND i64 with.
5349	* @param i64 The value to AND *pi64 with.
5350	*
5351	* @remarks x86: Requires a Pentium or later.
5352	*/
5353	DECLINLINE(void) ASMAtomicUoAndS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
5354	{
5355	ASMAtomicUoAndU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64);
5356	}
5357
5358
5359	/**
5360	* Atomically XOR an unsigned 32-bit value, unordered but interrupt safe.
5361	*
5362	* @param pu32 Pointer to the variable to XOR @a u32 with.
5363	* @param u32 The value to OR @a *pu32 with.
5364	*
5365	* @remarks x86: Requires a 386 or later.
5366	*/
5367	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
5368	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicUoXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
5369	#else
5370	DECLINLINE(void) ASMAtomicUoXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5371	{
5372	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5373	# if RT_INLINE_ASM_GNU_STYLE
5374	__asm__ __volatile__("xorl %1, %0\n\t"
5375	: "=m" (*pu32)
5376	: "ir" (u32)
5377	, "m" (*pu32)
5378	: "cc");
5379	# else
5380	__asm
5381	{
5382	mov eax, [u32]
5383	# ifdef RT_ARCH_AMD64
5384	mov rdx, [pu32]
5385	xor [rdx], eax
5386	# else
5387	mov edx, [pu32]
5388	xor [edx], eax
5389	# endif
5390	}
5391	# endif
5392
5393	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5394	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoXorU32, pu32, NO_BARRIER,
5395	"eor %w[uNew], %w[uNew], %w[uVal]\n\t",
5396	"eor %[uNew], %[uNew], %[uVal]\n\t",
5397	[uVal] "r" (u32));
5398
5399	# else
5400	# error "Port me"
5401	# endif
5402	}
5403	#endif
5404
5405
5406	/**
5407	* Atomically XOR an unsigned 32-bit value, unordered but interrupt safe,
5408	* extended version (for bitmap fallback).
5409	*
5410	* @returns Old value.
5411	* @param pu32 Pointer to the variable to XOR @a u32 with.
5412	* @param u32 The value to OR @a *pu32 with.
5413	*/
5414	DECLINLINE(uint32_t) ASMAtomicUoXorExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5415	{
5416	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5417	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicUoXorExU32, pu32, NO_BARRIER,
5418	"eor %w[uNew], %w[uOld], %w[uVal]\n\t",
5419	"eor %[uNew], %[uOld], %[uVal]\n\t",
5420	[uVal] "r" (u32));
5421	return u32OldRet;
5422
5423	#else
5424	return ASMAtomicXorExU32(pu32, u32); /* (we have no unordered cmpxchg primitive atm.) */
5425	#endif
5426	}
5427
5428
5429	/**
5430	* Atomically XOR a signed 32-bit value, unordered.
5431	*
5432	* @param pi32 Pointer to the variable to XOR @a u32 with.
5433	* @param i32 The value to XOR @a *pu32 with.
5434	*
5435	* @remarks x86: Requires a 386 or later.
5436	*/
5437	DECLINLINE(void) ASMAtomicUoXorS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
5438	{
5439	ASMAtomicUoXorU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
5440	}
5441
5442
5443	/**
5444	* Atomically increment an unsigned 32-bit value, unordered.
5445	*
5446	* @returns the new value.
5447	* @param pu32 Pointer to the variable to increment.
5448	*
5449	* @remarks x86: Requires a 486 or later.
5450	*/
5451	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
5452	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicUoIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO;
5453	#else
5454	DECLINLINE(uint32_t) ASMAtomicUoIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF
5455	{
5456	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5457	uint32_t u32;
5458	# if RT_INLINE_ASM_GNU_STYLE
5459	__asm__ __volatile__("xaddl %0, %1\n\t"
5460	: "=r" (u32)
5461	, "=m" (*pu32)
5462	: "0" (1)
5463	, "m" (*pu32)
5464	: "memory" /** @todo why 'memory'? */
5465	, "cc");
5466	return u32 + 1;
5467	# else
5468	__asm
5469	{
5470	mov eax, 1
5471	# ifdef RT_ARCH_AMD64
5472	mov rdx, [pu32]
5473	xadd [rdx], eax
5474	# else
5475	mov edx, [pu32]
5476	xadd [edx], eax
5477	# endif
5478	mov u32, eax
5479	}
5480	return u32 + 1;
5481	# endif
5482
5483	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5484	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoIncU32, pu32, NO_BARRIER,
5485	"add %w[uNew], %w[uNew], #1\n\t",
5486	"add %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */,
5487	"X" (0) /* dummy */);
5488	return u32NewRet;
5489
5490	# else
5491	# error "Port me"
5492	# endif
5493	}
5494	#endif
5495
5496
5497	/**
5498	* Atomically decrement an unsigned 32-bit value, unordered.
5499	*
5500	* @returns the new value.
5501	* @param pu32 Pointer to the variable to decrement.
5502	*
5503	* @remarks x86: Requires a 486 or later.
5504	*/
5505	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
5506	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicUoDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO;
5507	#else
5508	DECLINLINE(uint32_t) ASMAtomicUoDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF
5509	{
5510	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5511	uint32_t u32;
5512	# if RT_INLINE_ASM_GNU_STYLE
5513	__asm__ __volatile__("lock; xaddl %0, %1\n\t"
5514	: "=r" (u32)
5515	, "=m" (*pu32)
5516	: "0" (-1)
5517	, "m" (*pu32)
5518	: "memory"
5519	, "cc");
5520	return u32 - 1;
5521	# else
5522	__asm
5523	{
5524	mov eax, -1
5525	# ifdef RT_ARCH_AMD64
5526	mov rdx, [pu32]
5527	xadd [rdx], eax
5528	# else
5529	mov edx, [pu32]
5530	xadd [edx], eax
5531	# endif
5532	mov u32, eax
5533	}
5534	return u32 - 1;
5535	# endif
5536
5537	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5538	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoDecU32, pu32, NO_BARRIER,
5539	"sub %w[uNew], %w[uNew], #1\n\t",
5540	"sub %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */,
5541	"X" (0) /* dummy */);
5542	return u32NewRet;
5543
5544	# else
5545	# error "Port me"
5546	# endif
5547	}
5548	#endif
5549
5550
5551	/** @def RT_ASM_PAGE_SIZE
5552	* We try avoid dragging in iprt/param.h here.
5553	* @internal
5554	*/
5555	#if defined(RT_ARCH_SPARC64)
5556	# define RT_ASM_PAGE_SIZE 0x2000
5557	# if defined(PAGE_SIZE) && !defined(NT_INCLUDED)
5558	# if PAGE_SIZE != 0x2000
5559	# error "PAGE_SIZE is not 0x2000!"
5560	# endif
5561	# endif
5562	#elif defined(RT_ARCH_ARM64)
5563	# define RT_ASM_PAGE_SIZE 0x4000
5564	# if defined(PAGE_SIZE) && !defined(NT_INCLUDED) && !defined(_MACH_ARM_VM_PARAM_H_)
5565	# if PAGE_SIZE != 0x4000
5566	# error "PAGE_SIZE is not 0x4000!"
5567	# endif
5568	# endif
5569	#else
5570	# define RT_ASM_PAGE_SIZE 0x1000
5571	# if defined(PAGE_SIZE) && !defined(NT_INCLUDED)
5572	# if PAGE_SIZE != 0x1000
5573	# error "PAGE_SIZE is not 0x1000!"
5574	# endif
5575	# endif
5576	#endif
5577
5578	/**
5579	* Zeros a 4K memory page.
5580	*
5581	* @param pv Pointer to the memory block. This must be page aligned.
5582	*/
5583	#if (RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN) \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
5584	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMMemZeroPage(volatile void RT_FAR *pv) RT_NOTHROW_PROTO;
5585	# else
5586	DECLINLINE(void) ASMMemZeroPage(volatile void RT_FAR *pv) RT_NOTHROW_DEF
5587	{
5588	# if RT_INLINE_ASM_USES_INTRIN
5589	# ifdef RT_ARCH_AMD64
5590	__stosq((unsigned __int64 *)pv, 0, RT_ASM_PAGE_SIZE / 8);
5591	# else
5592	__stosd((unsigned long *)pv, 0, RT_ASM_PAGE_SIZE / 4);
5593	# endif
5594
5595	# elif RT_INLINE_ASM_GNU_STYLE
5596	RTCCUINTREG uDummy;
5597	# ifdef RT_ARCH_AMD64
5598	__asm__ __volatile__("rep stosq"
5599	: "=D" (pv),
5600	"=c" (uDummy)
5601	: "0" (pv),
5602	"c" (RT_ASM_PAGE_SIZE >> 3),
5603	"a" (0)
5604	: "memory");
5605	# else
5606	__asm__ __volatile__("rep stosl"
5607	: "=D" (pv),
5608	"=c" (uDummy)
5609	: "0" (pv),
5610	"c" (RT_ASM_PAGE_SIZE >> 2),
5611	"a" (0)
5612	: "memory");
5613	# endif
5614	# else
5615	__asm
5616	{
5617	# ifdef RT_ARCH_AMD64
5618	xor rax, rax
5619	mov ecx, 0200h
5620	mov rdi, [pv]
5621	rep stosq
5622	# else
5623	xor eax, eax
5624	mov ecx, 0400h
5625	mov edi, [pv]
5626	rep stosd
5627	# endif
5628	}
5629	# endif
5630	}
5631	# endif
5632
5633
5634	/**
5635	* Zeros a memory block with a 32-bit aligned size.
5636	*
5637	* @param pv Pointer to the memory block.
5638	* @param cb Number of bytes in the block. This MUST be aligned on 32-bit!
5639	*/
5640	#if (RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN) \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
5641	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMMemZero32(volatile void RT_FAR *pv, size_t cb) RT_NOTHROW_PROTO;
5642	#else
5643	DECLINLINE(void) ASMMemZero32(volatile void RT_FAR *pv, size_t cb) RT_NOTHROW_DEF
5644	{
5645	# if RT_INLINE_ASM_USES_INTRIN
5646	# ifdef RT_ARCH_AMD64
5647	if (!(cb & 7))
5648	__stosq((unsigned __int64 RT_FAR *)pv, 0, cb / 8);
5649	else
5650	# endif
5651	__stosd((unsigned long RT_FAR *)pv, 0, cb / 4);
5652
5653	# elif RT_INLINE_ASM_GNU_STYLE
5654	__asm__ __volatile__("rep stosl"
5655	: "=D" (pv),
5656	"=c" (cb)
5657	: "0" (pv),
5658	"1" (cb >> 2),
5659	"a" (0)
5660	: "memory");
5661	# else
5662	__asm
5663	{
5664	xor eax, eax
5665	# ifdef RT_ARCH_AMD64
5666	mov rcx, [cb]
5667	shr rcx, 2
5668	mov rdi, [pv]
5669	# else
5670	mov ecx, [cb]
5671	shr ecx, 2
5672	mov edi, [pv]
5673	# endif
5674	rep stosd
5675	}
5676	# endif
5677	}
5678	#endif
5679
5680
5681	/**
5682	* Fills a memory block with a 32-bit aligned size.
5683	*
5684	* @param pv Pointer to the memory block.
5685	* @param cb Number of bytes in the block. This MUST be aligned on 32-bit!
5686	* @param u32 The value to fill with.
5687	*/
5688	#if (RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN) \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
5689	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMMemFill32(volatile void RT_FAR *pv, size_t cb, uint32_t u32) RT_NOTHROW_PROTO;
5690	#else
5691	DECLINLINE(void) ASMMemFill32(volatile void RT_FAR *pv, size_t cb, uint32_t u32) RT_NOTHROW_DEF
5692	{
5693	# if RT_INLINE_ASM_USES_INTRIN
5694	# ifdef RT_ARCH_AMD64
5695	if (!(cb & 7))
5696	__stosq((unsigned __int64 RT_FAR *)pv, RT_MAKE_U64(u32, u32), cb / 8);
5697	else
5698	# endif
5699	__stosd((unsigned long RT_FAR *)pv, u32, cb / 4);
5700
5701	# elif RT_INLINE_ASM_GNU_STYLE
5702	__asm__ __volatile__("rep stosl"
5703	: "=D" (pv),
5704	"=c" (cb)
5705	: "0" (pv),
5706	"1" (cb >> 2),
5707	"a" (u32)
5708	: "memory");
5709	# else
5710	__asm
5711	{
5712	# ifdef RT_ARCH_AMD64
5713	mov rcx, [cb]
5714	shr rcx, 2
5715	mov rdi, [pv]
5716	# else
5717	mov ecx, [cb]
5718	shr ecx, 2
5719	mov edi, [pv]
5720	# endif
5721	mov eax, [u32]
5722	rep stosd
5723	}
5724	# endif
5725	}
5726	#endif
5727
5728
5729	/**
5730	* Checks if a memory block is all zeros.
5731	*
5732	* @returns Pointer to the first non-zero byte.
5733	* @returns NULL if all zero.
5734	*
5735	* @param pv Pointer to the memory block.
5736	* @param cb Number of bytes in the block.
5737	*/
5738	#if !defined(RDESKTOP) && (!defined(RT_OS_LINUX) \|\| !defined(__KERNEL__))
5739	DECLASM(void RT_FAR ) ASMMemFirstNonZero(void const RT_FAR pv, size_t cb) RT_NOTHROW_PROTO;
5740	#else
5741	DECLINLINE(void RT_FAR ) ASMMemFirstNonZero(void const RT_FAR pv, size_t cb) RT_NOTHROW_DEF
5742	{
5743	/** @todo replace with ASMMemFirstNonZero-generic.cpp in kernel modules. */
5744	uint8_t const pb = (uint8_t const RT_FAR )pv;
5745	for (; cb; cb--, pb++)
5746	if (RT_LIKELY(*pb == 0))
5747	{ /* likely */ }
5748	else
5749	return (void RT_FAR *)pb;
5750	return NULL;
5751	}
5752	#endif
5753
5754
5755	/**
5756	* Checks if a memory block is all zeros.
5757	*
5758	* @returns true if zero, false if not.
5759	*
5760	* @param pv Pointer to the memory block.
5761	* @param cb Number of bytes in the block.
5762	*
5763	* @sa ASMMemFirstNonZero
5764	*/
5765	DECLINLINE(bool) ASMMemIsZero(void const RT_FAR *pv, size_t cb) RT_NOTHROW_DEF
5766	{
5767	return ASMMemFirstNonZero(pv, cb) == NULL;
5768	}
5769
5770
5771	/**
5772	* Checks if a memory page is all zeros.
5773	*
5774	* @returns true / false.
5775	*
5776	* @param pvPage Pointer to the page. Must be aligned on 16 byte
5777	* boundary
5778	*/
5779	DECLINLINE(bool) ASMMemIsZeroPage(void const RT_FAR *pvPage) RT_NOTHROW_DEF
5780	{
5781	# if 0 /RT_INLINE_ASM_GNU_STYLE - this is actually slower... /
5782	union { RTCCUINTREG r; bool f; } uAX;
5783	RTCCUINTREG xCX, xDI;
5784	Assert(!((uintptr_t)pvPage & 15));
5785	__asm__ __volatile__("repe; "
5786	# ifdef RT_ARCH_AMD64
5787	"scasq\n\t"
5788	# else
5789	"scasl\n\t"
5790	# endif
5791	"setnc %%al\n\t"
5792	: "=&c" (xCX)
5793	, "=&D" (xDI)
5794	, "=&a" (uAX.r)
5795	: "mr" (pvPage)
5796	# ifdef RT_ARCH_AMD64
5797	, "0" (RT_ASM_PAGE_SIZE/8)
5798	# else
5799	, "0" (RT_ASM_PAGE_SIZE/4)
5800	# endif
5801	, "1" (pvPage)
5802	, "2" (0)
5803	: "cc");
5804	return uAX.f;
5805	# else
5806	uintptr_t const RT_FAR puPtr = (uintptr_t const RT_FAR )pvPage;
5807	size_t cLeft = RT_ASM_PAGE_SIZE / sizeof(uintptr_t) / 8;
5808	Assert(!((uintptr_t)pvPage & 15));
5809	for (;;)
5810	{
5811	if (puPtr[0]) return false;
5812	if (puPtr[4]) return false;
5813
5814	if (puPtr[2]) return false;
5815	if (puPtr[6]) return false;
5816
5817	if (puPtr[1]) return false;
5818	if (puPtr[5]) return false;
5819
5820	if (puPtr[3]) return false;
5821	if (puPtr[7]) return false;
5822
5823	if (!--cLeft)
5824	return true;
5825	puPtr += 8;
5826	}
5827	# endif
5828	}
5829
5830
5831	/**
5832	* Checks if a memory block is filled with the specified byte, returning the
5833	* first mismatch.
5834	*
5835	* This is sort of an inverted memchr.
5836	*
5837	* @returns Pointer to the byte which doesn't equal u8.
5838	* @returns NULL if all equal to u8.
5839	*
5840	* @param pv Pointer to the memory block.
5841	* @param cb Number of bytes in the block.
5842	* @param u8 The value it's supposed to be filled with.
5843	*
5844	* @remarks No alignment requirements.
5845	*/
5846	#if (!defined(RT_OS_LINUX) \|\| !defined(__KERNEL__)) \
5847	&& (!defined(RT_OS_FREEBSD) \|\| !defined(_KERNEL))
5848	DECLASM(void ) ASMMemFirstMismatchingU8(void const RT_FAR pv, size_t cb, uint8_t u8) RT_NOTHROW_PROTO;
5849	#else
5850	DECLINLINE(void ) ASMMemFirstMismatchingU8(void const RT_FAR pv, size_t cb, uint8_t u8) RT_NOTHROW_DEF
5851	{
5852	/** @todo replace with ASMMemFirstMismatchingU8-generic.cpp in kernel modules. */
5853	uint8_t const pb = (uint8_t const RT_FAR )pv;
5854	for (; cb; cb--, pb++)
5855	if (RT_LIKELY(*pb == u8))
5856	{ /* likely */ }
5857	else
5858	return (void *)pb;
5859	return NULL;
5860	}
5861	#endif
5862
5863
5864	/**
5865	* Checks if a memory block is filled with the specified byte.
5866	*
5867	* @returns true if all matching, false if not.
5868	*
5869	* @param pv Pointer to the memory block.
5870	* @param cb Number of bytes in the block.
5871	* @param u8 The value it's supposed to be filled with.
5872	*
5873	* @remarks No alignment requirements.
5874	*/
5875	DECLINLINE(bool) ASMMemIsAllU8(void const RT_FAR *pv, size_t cb, uint8_t u8) RT_NOTHROW_DEF
5876	{
5877	return ASMMemFirstMismatchingU8(pv, cb, u8) == NULL;
5878	}
5879
5880
5881	/**
5882	* Checks if a memory block is filled with the specified 32-bit value.
5883	*
5884	* This is a sort of inverted memchr.
5885	*
5886	* @returns Pointer to the first value which doesn't equal u32.
5887	* @returns NULL if all equal to u32.
5888	*
5889	* @param pv Pointer to the memory block.
5890	* @param cb Number of bytes in the block. This MUST be aligned on 32-bit!
5891	* @param u32 The value it's supposed to be filled with.
5892	*/
5893	DECLINLINE(uint32_t RT_FAR ) ASMMemFirstMismatchingU32(void const RT_FAR pv, size_t cb, uint32_t u32) RT_NOTHROW_DEF
5894	{
5895	/** @todo rewrite this in inline assembly? */
5896	uint32_t const RT_FAR pu32 = (uint32_t const RT_FAR )pv;
5897	for (; cb; cb -= 4, pu32++)
5898	if (RT_LIKELY(*pu32 == u32))
5899	{ /* likely */ }
5900	else
5901	return (uint32_t RT_FAR *)pu32;
5902	return NULL;
5903	}
5904
5905
5906	/**
5907	* Probes a byte pointer for read access.
5908	*
5909	* While the function will not fault if the byte is not read accessible,
5910	* the idea is to do this in a safe place like before acquiring locks
5911	* and such like.
5912	*
5913	* Also, this functions guarantees that an eager compiler is not going
5914	* to optimize the probing away.
5915	*
5916	* @param pvByte Pointer to the byte.
5917	*/
5918	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
5919	RT_ASM_DECL_PRAGMA_WATCOM(uint8_t) ASMProbeReadByte(const void RT_FAR *pvByte) RT_NOTHROW_PROTO;
5920	#else
5921	DECLINLINE(uint8_t) ASMProbeReadByte(const void RT_FAR *pvByte) RT_NOTHROW_DEF
5922	{
5923	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5924	uint8_t u8;
5925	# if RT_INLINE_ASM_GNU_STYLE
5926	__asm__ __volatile__("movb %1, %0\n\t"
5927	: "=q" (u8)
5928	: "m" ((const uint8_t )pvByte));
5929	# else
5930	__asm
5931	{
5932	# ifdef RT_ARCH_AMD64
5933	mov rax, [pvByte]
5934	mov al, [rax]
5935	# else
5936	mov eax, [pvByte]
5937	mov al, [eax]
5938	# endif
5939	mov [u8], al
5940	}
5941	# endif
5942	return u8;
5943
5944	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5945	uint32_t u32;
5946	__asm__ __volatile__(".Lstart_ASMProbeReadByte_%=:\n\t"
5947	# if defined(RT_ARCH_ARM64)
5948	"ldxrb %w[uDst], %[pMem]\n\t"
5949	# else
5950	"ldrexb %[uDst], %[pMem]\n\t"
5951	# endif
5952	: [uDst] "=&r" (u32)
5953	: [pMem] "Q" ((uint8_t const )pvByte));
5954	return (uint8_t)u32;
5955
5956	# else
5957	# error "Port me"
5958	# endif
5959	}
5960	#endif
5961
5962	/**
5963	* Probes a buffer for read access page by page.
5964	*
5965	* While the function will fault if the buffer is not fully read
5966	* accessible, the idea is to do this in a safe place like before
5967	* acquiring locks and such like.
5968	*
5969	* Also, this functions guarantees that an eager compiler is not going
5970	* to optimize the probing away.
5971	*
5972	* @param pvBuf Pointer to the buffer.
5973	* @param cbBuf The size of the buffer in bytes. Must be >= 1.
5974	*/
5975	DECLINLINE(void) ASMProbeReadBuffer(const void RT_FAR *pvBuf, size_t cbBuf) RT_NOTHROW_DEF
5976	{
5977	/** @todo verify that the compiler actually doesn't optimize this away. (intel & gcc) */
5978	/* the first byte */
5979	const uint8_t RT_FAR pu8 = (const uint8_t RT_FAR )pvBuf;
5980	ASMProbeReadByte(pu8);
5981
5982	/* the pages in between pages. */
5983	while (cbBuf > RT_ASM_PAGE_SIZE)
5984	{
5985	ASMProbeReadByte(pu8);
5986	cbBuf -= RT_ASM_PAGE_SIZE;
5987	pu8 += RT_ASM_PAGE_SIZE;
5988	}
5989
5990	/* the last byte */
5991	ASMProbeReadByte(pu8 + cbBuf - 1);
5992	}
5993
5994
5995	/**
5996	* Reverse the byte order of the given 16-bit integer.
5997	*
5998	* @returns Revert
5999	* @param u16 16-bit integer value.
6000	*/
6001	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6002	RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMByteSwapU16(uint16_t u16) RT_NOTHROW_PROTO;
6003	#else
6004	DECLINLINE(uint16_t) ASMByteSwapU16(uint16_t u16) RT_NOTHROW_DEF
6005	{
6006	# if RT_INLINE_ASM_USES_INTRIN
6007	return _byteswap_ushort(u16);
6008
6009	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6010	# if RT_INLINE_ASM_GNU_STYLE
6011	__asm__ ("rorw $8, %0" : "=r" (u16) : "0" (u16) : "cc");
6012	# else
6013	_asm
6014	{
6015	mov ax, [u16]
6016	ror ax, 8
6017	mov [u16], ax
6018	}
6019	# endif
6020	return u16;
6021
6022	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6023	uint32_t u32Ret;
6024	__asm__ __volatile__(
6025	# if defined(RT_ARCH_ARM64)
6026	"rev16 %w[uRet], %w[uVal]\n\t"
6027	# else
6028	"rev16 %[uRet], %[uVal]\n\t"
6029	# endif
6030	: [uRet] "=r" (u32Ret)
6031	: [uVal] "r" (u16));
6032	return (uint16_t)u32Ret;
6033
6034	# else
6035	# error "Port me"
6036	# endif
6037	}
6038	#endif
6039
6040
6041	/**
6042	* Reverse the byte order of the given 32-bit integer.
6043	*
6044	* @returns Revert
6045	* @param u32 32-bit integer value.
6046	*/
6047	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6048	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMByteSwapU32(uint32_t u32) RT_NOTHROW_PROTO;
6049	#else
6050	DECLINLINE(uint32_t) ASMByteSwapU32(uint32_t u32) RT_NOTHROW_DEF
6051	{
6052	# if RT_INLINE_ASM_USES_INTRIN
6053	return _byteswap_ulong(u32);
6054
6055	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6056	# if RT_INLINE_ASM_GNU_STYLE
6057	__asm__ ("bswapl %0" : "=r" (u32) : "0" (u32));
6058	# else
6059	_asm
6060	{
6061	mov eax, [u32]
6062	bswap eax
6063	mov [u32], eax
6064	}
6065	# endif
6066	return u32;
6067
6068	# elif defined(RT_ARCH_ARM64)
6069	uint64_t u64Ret;
6070	__asm__ __volatile__("rev32 %[uRet], %[uVal]\n\t"
6071	: [uRet] "=r" (u64Ret)
6072	: [uVal] "r" ((uint64_t)u32));
6073	return (uint32_t)u64Ret;
6074
6075	# elif defined(RT_ARCH_ARM32)
6076	__asm__ __volatile__("rev %[uRet], %[uVal]\n\t"
6077	: [uRet] "=r" (u32)
6078	: [uVal] "[uRet]" (u32));
6079	return u32;
6080
6081	# else
6082	# error "Port me"
6083	# endif
6084	}
6085	#endif
6086
6087
6088	/**
6089	* Reverse the byte order of the given 64-bit integer.
6090	*
6091	* @returns Revert
6092	* @param u64 64-bit integer value.
6093	*/
6094	DECLINLINE(uint64_t) ASMByteSwapU64(uint64_t u64) RT_NOTHROW_DEF
6095	{
6096	#if defined(RT_ARCH_AMD64) && RT_INLINE_ASM_USES_INTRIN
6097	return _byteswap_uint64(u64);
6098
6099	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
6100	__asm__ ("bswapq %0" : "=r" (u64) : "0" (u64));
6101	return u64;
6102
6103	# elif defined(RT_ARCH_ARM64)
6104	__asm__ __volatile__("rev %[uRet], %[uVal]\n\t"
6105	: [uRet] "=r" (u64)
6106	: [uVal] "[uRet]" (u64));
6107	return u64;
6108
6109	#else
6110	return (uint64_t)ASMByteSwapU32((uint32_t)u64) << 32
6111	\| (uint64_t)ASMByteSwapU32((uint32_t)(u64 >> 32));
6112	#endif
6113	}
6114
6115
6116
6117	/** @defgroup grp_inline_bits Bit Operations
6118	* @{
6119	*/
6120
6121
6122	/**
6123	* Sets a bit in a bitmap.
6124	*
6125	* @param pvBitmap Pointer to the bitmap (little endian). This should be
6126	* 32-bit aligned.
6127	* @param iBit The bit to set.
6128	*
6129	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
6130	* However, doing so will yield better performance as well as avoiding
6131	* traps accessing the last bits in the bitmap.
6132	*/
6133	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6134	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6135	#else
6136	DECLINLINE(void) ASMBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6137	{
6138	# if RT_INLINE_ASM_USES_INTRIN
6139	_bittestandset((long RT_FAR *)pvBitmap, iBit);
6140
6141	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6142	# if RT_INLINE_ASM_GNU_STYLE
6143	__asm__ __volatile__("btsl %1, %0"
6144	: "=m" ((volatile long RT_FAR )pvBitmap)
6145	: "Ir" (iBit)
6146	, "m" ((volatile long RT_FAR )pvBitmap)
6147	: "memory"
6148	, "cc");
6149	# else
6150	__asm
6151	{
6152	# ifdef RT_ARCH_AMD64
6153	mov rax, [pvBitmap]
6154	mov edx, [iBit]
6155	bts [rax], edx
6156	# else
6157	mov eax, [pvBitmap]
6158	mov edx, [iBit]
6159	bts [eax], edx
6160	# endif
6161	}
6162	# endif
6163
6164	# else
6165	int32_t offBitmap = iBit / 32;
6166	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
6167	ASMAtomicUoOrU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31)));
6168	# endif
6169	}
6170	#endif
6171
6172
6173	/**
6174	* Atomically sets a bit in a bitmap, ordered.
6175	*
6176	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
6177	* aligned, otherwise the memory access isn't atomic!
6178	* @param iBit The bit to set.
6179	*
6180	* @remarks x86: Requires a 386 or later.
6181	*/
6182	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6183	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6184	#else
6185	DECLINLINE(void) ASMAtomicBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6186	{
6187	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
6188	# if RT_INLINE_ASM_USES_INTRIN
6189	_interlockedbittestandset((long RT_FAR *)pvBitmap, iBit);
6190	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6191	# if RT_INLINE_ASM_GNU_STYLE
6192	__asm__ __volatile__("lock; btsl %1, %0"
6193	: "=m" ((volatile long )pvBitmap)
6194	: "Ir" (iBit)
6195	, "m" ((volatile long )pvBitmap)
6196	: "memory"
6197	, "cc");
6198	# else
6199	__asm
6200	{
6201	# ifdef RT_ARCH_AMD64
6202	mov rax, [pvBitmap]
6203	mov edx, [iBit]
6204	lock bts [rax], edx
6205	# else
6206	mov eax, [pvBitmap]
6207	mov edx, [iBit]
6208	lock bts [eax], edx
6209	# endif
6210	}
6211	# endif
6212
6213	# else
6214	ASMAtomicOrU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(RT_BIT_32(iBit & 31)));
6215	# endif
6216	}
6217	#endif
6218
6219
6220	/**
6221	* Clears a bit in a bitmap.
6222	*
6223	* @param pvBitmap Pointer to the bitmap (little endian).
6224	* @param iBit The bit to clear.
6225	*
6226	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
6227	* However, doing so will yield better performance as well as avoiding
6228	* traps accessing the last bits in the bitmap.
6229	*/
6230	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6231	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6232	#else
6233	DECLINLINE(void) ASMBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6234	{
6235	# if RT_INLINE_ASM_USES_INTRIN
6236	_bittestandreset((long RT_FAR *)pvBitmap, iBit);
6237
6238	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6239	# if RT_INLINE_ASM_GNU_STYLE
6240	__asm__ __volatile__("btrl %1, %0"
6241	: "=m" ((volatile long RT_FAR )pvBitmap)
6242	: "Ir" (iBit)
6243	, "m" ((volatile long RT_FAR )pvBitmap)
6244	: "memory"
6245	, "cc");
6246	# else
6247	__asm
6248	{
6249	# ifdef RT_ARCH_AMD64
6250	mov rax, [pvBitmap]
6251	mov edx, [iBit]
6252	btr [rax], edx
6253	# else
6254	mov eax, [pvBitmap]
6255	mov edx, [iBit]
6256	btr [eax], edx
6257	# endif
6258	}
6259	# endif
6260
6261	# else
6262	int32_t offBitmap = iBit / 32;
6263	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
6264	ASMAtomicUoAndU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(~RT_BIT_32(iBit & 31)));
6265	# endif
6266	}
6267	#endif
6268
6269
6270	/**
6271	* Atomically clears a bit in a bitmap, ordered.
6272	*
6273	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
6274	* aligned, otherwise the memory access isn't atomic!
6275	* @param iBit The bit to toggle set.
6276	*
6277	* @remarks No memory barrier, take care on smp.
6278	* @remarks x86: Requires a 386 or later.
6279	*/
6280	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
6281	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6282	#else
6283	DECLINLINE(void) ASMAtomicBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6284	{
6285	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
6286	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6287	# if RT_INLINE_ASM_GNU_STYLE
6288	__asm__ __volatile__("lock; btrl %1, %0"
6289	: "=m" ((volatile long RT_FAR )pvBitmap)
6290	: "Ir" (iBit)
6291	, "m" ((volatile long RT_FAR )pvBitmap)
6292	: "memory"
6293	, "cc");
6294	# else
6295	__asm
6296	{
6297	# ifdef RT_ARCH_AMD64
6298	mov rax, [pvBitmap]
6299	mov edx, [iBit]
6300	lock btr [rax], edx
6301	# else
6302	mov eax, [pvBitmap]
6303	mov edx, [iBit]
6304	lock btr [eax], edx
6305	# endif
6306	}
6307	# endif
6308	# else
6309	ASMAtomicAndU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(~RT_BIT_32(iBit & 31)));
6310	# endif
6311	}
6312	#endif
6313
6314
6315	/**
6316	* Toggles a bit in a bitmap.
6317	*
6318	* @param pvBitmap Pointer to the bitmap (little endian).
6319	* @param iBit The bit to toggle.
6320	*
6321	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
6322	* However, doing so will yield better performance as well as avoiding
6323	* traps accessing the last bits in the bitmap.
6324	*/
6325	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6326	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6327	#else
6328	DECLINLINE(void) ASMBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6329	{
6330	# if RT_INLINE_ASM_USES_INTRIN
6331	_bittestandcomplement((long RT_FAR *)pvBitmap, iBit);
6332	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6333	# if RT_INLINE_ASM_GNU_STYLE
6334	__asm__ __volatile__("btcl %1, %0"
6335	: "=m" ((volatile long )pvBitmap)
6336	: "Ir" (iBit)
6337	, "m" ((volatile long )pvBitmap)
6338	: "memory"
6339	, "cc");
6340	# else
6341	__asm
6342	{
6343	# ifdef RT_ARCH_AMD64
6344	mov rax, [pvBitmap]
6345	mov edx, [iBit]
6346	btc [rax], edx
6347	# else
6348	mov eax, [pvBitmap]
6349	mov edx, [iBit]
6350	btc [eax], edx
6351	# endif
6352	}
6353	# endif
6354	# else
6355	int32_t offBitmap = iBit / 32;
6356	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
6357	ASMAtomicUoXorU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31)));
6358	# endif
6359	}
6360	#endif
6361
6362
6363	/**
6364	* Atomically toggles a bit in a bitmap, ordered.
6365	*
6366	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
6367	* aligned, otherwise the memory access isn't atomic!
6368	* @param iBit The bit to test and set.
6369	*
6370	* @remarks x86: Requires a 386 or later.
6371	*/
6372	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
6373	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6374	#else
6375	DECLINLINE(void) ASMAtomicBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6376	{
6377	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
6378	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6379	# if RT_INLINE_ASM_GNU_STYLE
6380	__asm__ __volatile__("lock; btcl %1, %0"
6381	: "=m" ((volatile long RT_FAR )pvBitmap)
6382	: "Ir" (iBit)
6383	, "m" ((volatile long RT_FAR )pvBitmap)
6384	: "memory"
6385	, "cc");
6386	# else
6387	__asm
6388	{
6389	# ifdef RT_ARCH_AMD64
6390	mov rax, [pvBitmap]
6391	mov edx, [iBit]
6392	lock btc [rax], edx
6393	# else
6394	mov eax, [pvBitmap]
6395	mov edx, [iBit]
6396	lock btc [eax], edx
6397	# endif
6398	}
6399	# endif
6400	# else
6401	ASMAtomicXorU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(RT_BIT_32(iBit & 31)));
6402	# endif
6403	}
6404	#endif
6405
6406
6407	/**
6408	* Tests and sets a bit in a bitmap.
6409	*
6410	* @returns true if the bit was set.
6411	* @returns false if the bit was clear.
6412	*
6413	* @param pvBitmap Pointer to the bitmap (little endian).
6414	* @param iBit The bit to test and set.
6415	*
6416	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
6417	* However, doing so will yield better performance as well as avoiding
6418	* traps accessing the last bits in the bitmap.
6419	*/
6420	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6421	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6422	#else
6423	DECLINLINE(bool) ASMBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6424	{
6425	union { bool f; uint32_t u32; uint8_t u8; } rc;
6426	# if RT_INLINE_ASM_USES_INTRIN
6427	rc.u8 = _bittestandset((long RT_FAR *)pvBitmap, iBit);
6428
6429	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6430	# if RT_INLINE_ASM_GNU_STYLE
6431	__asm__ __volatile__("btsl %2, %1\n\t"
6432	"setc %b0\n\t"
6433	"andl $1, %0\n\t"
6434	: "=q" (rc.u32)
6435	, "=m" ((volatile long RT_FAR )pvBitmap)
6436	: "Ir" (iBit)
6437	, "m" ((volatile long RT_FAR )pvBitmap)
6438	: "memory"
6439	, "cc");
6440	# else
6441	__asm
6442	{
6443	mov edx, [iBit]
6444	# ifdef RT_ARCH_AMD64
6445	mov rax, [pvBitmap]
6446	bts [rax], edx
6447	# else
6448	mov eax, [pvBitmap]
6449	bts [eax], edx
6450	# endif
6451	setc al
6452	and eax, 1
6453	mov [rc.u32], eax
6454	}
6455	# endif
6456
6457	# else
6458	int32_t offBitmap = iBit / 32;
6459	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
6460	rc.u32 = RT_LE2H_U32(ASMAtomicUoOrExU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31))))
6461	>> (iBit & 31);
6462	rc.u32 &= 1;
6463	# endif
6464	return rc.f;
6465	}
6466	#endif
6467
6468
6469	/**
6470	* Atomically tests and sets a bit in a bitmap, ordered.
6471	*
6472	* @returns true if the bit was set.
6473	* @returns false if the bit was clear.
6474	*
6475	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
6476	* aligned, otherwise the memory access isn't atomic!
6477	* @param iBit The bit to set.
6478	*
6479	* @remarks x86: Requires a 386 or later.
6480	*/
6481	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6482	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6483	#else
6484	DECLINLINE(bool) ASMAtomicBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6485	{
6486	union { bool f; uint32_t u32; uint8_t u8; } rc;
6487	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
6488	# if RT_INLINE_ASM_USES_INTRIN
6489	rc.u8 = _interlockedbittestandset((long RT_FAR *)pvBitmap, iBit);
6490	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6491	# if RT_INLINE_ASM_GNU_STYLE
6492	__asm__ __volatile__("lock; btsl %2, %1\n\t"
6493	"setc %b0\n\t"
6494	"andl $1, %0\n\t"
6495	: "=q" (rc.u32)
6496	, "=m" ((volatile long RT_FAR )pvBitmap)
6497	: "Ir" (iBit)
6498	, "m" ((volatile long RT_FAR )pvBitmap)
6499	: "memory"
6500	, "cc");
6501	# else
6502	__asm
6503	{
6504	mov edx, [iBit]
6505	# ifdef RT_ARCH_AMD64
6506	mov rax, [pvBitmap]
6507	lock bts [rax], edx
6508	# else
6509	mov eax, [pvBitmap]
6510	lock bts [eax], edx
6511	# endif
6512	setc al
6513	and eax, 1
6514	mov [rc.u32], eax
6515	}
6516	# endif
6517
6518	# else
6519	rc.u32 = RT_LE2H_U32(ASMAtomicOrExU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(RT_BIT_32(iBit & 31))))
6520	>> (iBit & 31);
6521	rc.u32 &= 1;
6522	# endif
6523	return rc.f;
6524	}
6525	#endif
6526
6527
6528	/**
6529	* Tests and clears a bit in a bitmap.
6530	*
6531	* @returns true if the bit was set.
6532	* @returns false if the bit was clear.
6533	*
6534	* @param pvBitmap Pointer to the bitmap (little endian).
6535	* @param iBit The bit to test and clear.
6536	*
6537	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
6538	* However, doing so will yield better performance as well as avoiding
6539	* traps accessing the last bits in the bitmap.
6540	*/
6541	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6542	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6543	#else
6544	DECLINLINE(bool) ASMBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6545	{
6546	union { bool f; uint32_t u32; uint8_t u8; } rc;
6547	# if RT_INLINE_ASM_USES_INTRIN
6548	rc.u8 = _bittestandreset((long RT_FAR *)pvBitmap, iBit);
6549
6550	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6551	# if RT_INLINE_ASM_GNU_STYLE
6552	__asm__ __volatile__("btrl %2, %1\n\t"
6553	"setc %b0\n\t"
6554	"andl $1, %0\n\t"
6555	: "=q" (rc.u32)
6556	, "=m" ((volatile long RT_FAR )pvBitmap)
6557	: "Ir" (iBit)
6558	, "m" ((volatile long RT_FAR )pvBitmap)
6559	: "memory"
6560	, "cc");
6561	# else
6562	__asm
6563	{
6564	mov edx, [iBit]
6565	# ifdef RT_ARCH_AMD64
6566	mov rax, [pvBitmap]
6567	btr [rax], edx
6568	# else
6569	mov eax, [pvBitmap]
6570	btr [eax], edx
6571	# endif
6572	setc al
6573	and eax, 1
6574	mov [rc.u32], eax
6575	}
6576	# endif
6577
6578	# else
6579	int32_t offBitmap = iBit / 32;
6580	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
6581	rc.u32 = RT_LE2H_U32(ASMAtomicUoAndExU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(~RT_BIT_32(iBit & 31))))
6582	>> (iBit & 31);
6583	rc.u32 &= 1;
6584	# endif
6585	return rc.f;
6586	}
6587	#endif
6588
6589
6590	/**
6591	* Atomically tests and clears a bit in a bitmap, ordered.
6592	*
6593	* @returns true if the bit was set.
6594	* @returns false if the bit was clear.
6595	*
6596	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
6597	* aligned, otherwise the memory access isn't atomic!
6598	* @param iBit The bit to test and clear.
6599	*
6600	* @remarks No memory barrier, take care on smp.
6601	* @remarks x86: Requires a 386 or later.
6602	*/
6603	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6604	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6605	#else
6606	DECLINLINE(bool) ASMAtomicBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6607	{
6608	union { bool f; uint32_t u32; uint8_t u8; } rc;
6609	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
6610	# if RT_INLINE_ASM_USES_INTRIN
6611	rc.u8 = _interlockedbittestandreset((long RT_FAR *)pvBitmap, iBit);
6612
6613	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6614	# if RT_INLINE_ASM_GNU_STYLE
6615	__asm__ __volatile__("lock; btrl %2, %1\n\t"
6616	"setc %b0\n\t"
6617	"andl $1, %0\n\t"
6618	: "=q" (rc.u32)
6619	, "=m" ((volatile long RT_FAR )pvBitmap)
6620	: "Ir" (iBit)
6621	, "m" ((volatile long RT_FAR )pvBitmap)
6622	: "memory"
6623	, "cc");
6624	# else
6625	__asm
6626	{
6627	mov edx, [iBit]
6628	# ifdef RT_ARCH_AMD64
6629	mov rax, [pvBitmap]
6630	lock btr [rax], edx
6631	# else
6632	mov eax, [pvBitmap]
6633	lock btr [eax], edx
6634	# endif
6635	setc al
6636	and eax, 1
6637	mov [rc.u32], eax
6638	}
6639	# endif
6640
6641	# else
6642	rc.u32 = RT_LE2H_U32(ASMAtomicAndExU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(~RT_BIT_32(iBit & 31))))
6643	>> (iBit & 31);
6644	rc.u32 &= 1;
6645	# endif
6646	return rc.f;
6647	}
6648	#endif
6649
6650
6651	/**
6652	* Tests and toggles a bit in a bitmap.
6653	*
6654	* @returns true if the bit was set.
6655	* @returns false if the bit was clear.
6656	*
6657	* @param pvBitmap Pointer to the bitmap (little endian).
6658	* @param iBit The bit to test and toggle.
6659	*
6660	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
6661	* However, doing so will yield better performance as well as avoiding
6662	* traps accessing the last bits in the bitmap.
6663	*/
6664	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6665	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6666	#else
6667	DECLINLINE(bool) ASMBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6668	{
6669	union { bool f; uint32_t u32; uint8_t u8; } rc;
6670	# if RT_INLINE_ASM_USES_INTRIN
6671	rc.u8 = _bittestandcomplement((long RT_FAR *)pvBitmap, iBit);
6672
6673	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6674	# if RT_INLINE_ASM_GNU_STYLE
6675	__asm__ __volatile__("btcl %2, %1\n\t"
6676	"setc %b0\n\t"
6677	"andl $1, %0\n\t"
6678	: "=q" (rc.u32)
6679	, "=m" ((volatile long RT_FAR )pvBitmap)
6680	: "Ir" (iBit)
6681	, "m" ((volatile long RT_FAR )pvBitmap)
6682	: "memory"
6683	, "cc");
6684	# else
6685	__asm
6686	{
6687	mov edx, [iBit]
6688	# ifdef RT_ARCH_AMD64
6689	mov rax, [pvBitmap]
6690	btc [rax], edx
6691	# else
6692	mov eax, [pvBitmap]
6693	btc [eax], edx
6694	# endif
6695	setc al
6696	and eax, 1
6697	mov [rc.u32], eax
6698	}
6699	# endif
6700
6701	# else
6702	int32_t offBitmap = iBit / 32;
6703	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
6704	rc.u32 = RT_LE2H_U32(ASMAtomicUoXorExU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31))))
6705	>> (iBit & 31);
6706	rc.u32 &= 1;
6707	# endif
6708	return rc.f;
6709	}
6710	#endif
6711
6712
6713	/**
6714	* Atomically tests and toggles a bit in a bitmap, ordered.
6715	*
6716	* @returns true if the bit was set.
6717	* @returns false if the bit was clear.
6718	*
6719	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
6720	* aligned, otherwise the memory access isn't atomic!
6721	* @param iBit The bit to test and toggle.
6722	*
6723	* @remarks x86: Requires a 386 or later.
6724	*/
6725	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
6726	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6727	#else
6728	DECLINLINE(bool) ASMAtomicBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6729	{
6730	union { bool f; uint32_t u32; uint8_t u8; } rc;
6731	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
6732	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6733	# if RT_INLINE_ASM_GNU_STYLE
6734	__asm__ __volatile__("lock; btcl %2, %1\n\t"
6735	"setc %b0\n\t"
6736	"andl $1, %0\n\t"
6737	: "=q" (rc.u32)
6738	, "=m" ((volatile long RT_FAR )pvBitmap)
6739	: "Ir" (iBit)
6740	, "m" ((volatile long RT_FAR )pvBitmap)
6741	: "memory"
6742	, "cc");
6743	# else
6744	__asm
6745	{
6746	mov edx, [iBit]
6747	# ifdef RT_ARCH_AMD64
6748	mov rax, [pvBitmap]
6749	lock btc [rax], edx
6750	# else
6751	mov eax, [pvBitmap]
6752	lock btc [eax], edx
6753	# endif
6754	setc al
6755	and eax, 1
6756	mov [rc.u32], eax
6757	}
6758	# endif
6759
6760	# else
6761	rc.u32 = RT_H2LE_U32(ASMAtomicXorExU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_LE2H_U32(RT_BIT_32(iBit & 31))))
6762	>> (iBit & 31);
6763	rc.u32 &= 1;
6764	# endif
6765	return rc.f;
6766	}
6767	#endif
6768
6769
6770	/**
6771	* Tests if a bit in a bitmap is set.
6772	*
6773	* @returns true if the bit is set.
6774	* @returns false if the bit is clear.
6775	*
6776	* @param pvBitmap Pointer to the bitmap (little endian).
6777	* @param iBit The bit to test.
6778	*
6779	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
6780	* However, doing so will yield better performance as well as avoiding
6781	* traps accessing the last bits in the bitmap.
6782	*/
6783	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6784	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTest(const volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6785	#else
6786	DECLINLINE(bool) ASMBitTest(const volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6787	{
6788	union { bool f; uint32_t u32; uint8_t u8; } rc;
6789	# if RT_INLINE_ASM_USES_INTRIN
6790	rc.u32 = _bittest((long *)pvBitmap, iBit);
6791
6792	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6793	# if RT_INLINE_ASM_GNU_STYLE
6794
6795	__asm__ __volatile__("btl %2, %1\n\t"
6796	"setc %b0\n\t"
6797	"andl $1, %0\n\t"
6798	: "=q" (rc.u32)
6799	: "m" ((const volatile long RT_FAR )pvBitmap)
6800	, "Ir" (iBit)
6801	: "memory"
6802	, "cc");
6803	# else
6804	__asm
6805	{
6806	mov edx, [iBit]
6807	# ifdef RT_ARCH_AMD64
6808	mov rax, [pvBitmap]
6809	bt [rax], edx
6810	# else
6811	mov eax, [pvBitmap]
6812	bt [eax], edx
6813	# endif
6814	setc al
6815	and eax, 1
6816	mov [rc.u32], eax
6817	}
6818	# endif
6819
6820	# else
6821	int32_t offBitmap = iBit / 32;
6822	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
6823	rc.u32 = RT_LE2H_U32(ASMAtomicUoReadU32(&((uint32_t volatile *)pvBitmap)[offBitmap])) >> (iBit & 31);
6824	rc.u32 &= 1;
6825	# endif
6826	return rc.f;
6827	}
6828	#endif
6829
6830
6831	/**
6832	* Clears a bit range within a bitmap.
6833	*
6834	* @param pvBitmap Pointer to the bitmap (little endian).
6835	* @param iBitStart The First bit to clear.
6836	* @param iBitEnd The first bit not to clear.
6837	*/
6838	DECLINLINE(void) ASMBitClearRange(volatile void RT_FAR *pvBitmap, size_t iBitStart, size_t iBitEnd) RT_NOTHROW_DEF
6839	{
6840	if (iBitStart < iBitEnd)
6841	{
6842	uint32_t volatile RT_FAR pu32 = (volatile uint32_t RT_FAR )pvBitmap + (iBitStart >> 5);
6843	size_t iStart = iBitStart & ~(size_t)31;
6844	size_t iEnd = iBitEnd & ~(size_t)31;
6845	if (iStart == iEnd)
6846	*pu32 &= RT_H2LE_U32(((UINT32_C(1) << (iBitStart & 31)) - 1) \| ~((UINT32_C(1) << (iBitEnd & 31)) - 1));
6847	else
6848	{
6849	/* bits in first dword. */
6850	if (iBitStart & 31)
6851	{
6852	*pu32 &= RT_H2LE_U32((UINT32_C(1) << (iBitStart & 31)) - 1);
6853	pu32++;
6854	iBitStart = iStart + 32;
6855	}
6856
6857	/* whole dwords. */
6858	if (iBitStart != iEnd)
6859	ASMMemZero32(pu32, (iEnd - iBitStart) >> 3);
6860
6861	/* bits in last dword. */
6862	if (iBitEnd & 31)
6863	{
6864	pu32 = (volatile uint32_t RT_FAR *)pvBitmap + (iBitEnd >> 5);
6865	*pu32 &= RT_H2LE_U32(~((UINT32_C(1) << (iBitEnd & 31)) - 1));
6866	}
6867	}
6868	}
6869	}
6870
6871
6872	/**
6873	* Sets a bit range within a bitmap.
6874	*
6875	* @param pvBitmap Pointer to the bitmap (little endian).
6876	* @param iBitStart The First bit to set.
6877	* @param iBitEnd The first bit not to set.
6878	*/
6879	DECLINLINE(void) ASMBitSetRange(volatile void RT_FAR *pvBitmap, size_t iBitStart, size_t iBitEnd) RT_NOTHROW_DEF
6880	{
6881	if (iBitStart < iBitEnd)
6882	{
6883	uint32_t volatile RT_FAR pu32 = (volatile uint32_t RT_FAR )pvBitmap + (iBitStart >> 5);
6884	size_t iStart = iBitStart & ~(size_t)31;
6885	size_t iEnd = iBitEnd & ~(size_t)31;
6886	if (iStart == iEnd)
6887	*pu32 \|= RT_H2LE_U32(((UINT32_C(1) << (iBitEnd - iBitStart)) - 1) << (iBitStart & 31));
6888	else
6889	{
6890	/* bits in first dword. */
6891	if (iBitStart & 31)
6892	{
6893	*pu32 \|= RT_H2LE_U32(~((UINT32_C(1) << (iBitStart & 31)) - 1));
6894	pu32++;
6895	iBitStart = iStart + 32;
6896	}
6897
6898	/* whole dword. */
6899	if (iBitStart != iEnd)
6900	ASMMemFill32(pu32, (iEnd - iBitStart) >> 3, ~UINT32_C(0));
6901
6902	/* bits in last dword. */
6903	if (iBitEnd & 31)
6904	{
6905	pu32 = (volatile uint32_t RT_FAR *)pvBitmap + (iBitEnd >> 5);
6906	*pu32 \|= RT_H2LE_U32((UINT32_C(1) << (iBitEnd & 31)) - 1);
6907	}
6908	}
6909	}
6910	}
6911
6912
6913	/**
6914	* Finds the first clear bit in a bitmap.
6915	*
6916	* @returns Index of the first zero bit.
6917	* @returns -1 if no clear bit was found.
6918	* @param pvBitmap Pointer to the bitmap (little endian).
6919	* @param cBits The number of bits in the bitmap. Multiple of 32.
6920	*/
6921	#if RT_INLINE_ASM_EXTERNAL \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
6922	DECLASM(int32_t) ASMBitFirstClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_PROTO;
6923	#else
6924	DECLINLINE(int32_t) ASMBitFirstClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_DEF
6925	{
6926	if (cBits)
6927	{
6928	int32_t iBit;
6929	# if RT_INLINE_ASM_GNU_STYLE
6930	RTCCUINTREG uEAX, uECX, uEDI;
6931	cBits = RT_ALIGN_32(cBits, 32);
6932	__asm__ __volatile__("repe; scasl\n\t"
6933	"je 1f\n\t"
6934	# ifdef RT_ARCH_AMD64
6935	"lea -4(%%rdi), %%rdi\n\t"
6936	"xorl (%%rdi), %%eax\n\t"
6937	"subq %5, %%rdi\n\t"
6938	# else
6939	"lea -4(%%edi), %%edi\n\t"
6940	"xorl (%%edi), %%eax\n\t"
6941	"subl %5, %%edi\n\t"
6942	# endif
6943	"shll $3, %%edi\n\t"
6944	"bsfl %%eax, %%edx\n\t"
6945	"addl %%edi, %%edx\n\t"
6946	"1:\t\n"
6947	: "=d" (iBit)
6948	, "=&c" (uECX)
6949	, "=&D" (uEDI)
6950	, "=&a" (uEAX)
6951	: "0" (0xffffffff)
6952	, "mr" (pvBitmap)
6953	, "1" (cBits >> 5)
6954	, "2" (pvBitmap)
6955	, "3" (0xffffffff)
6956	: "cc");
6957	# else
6958	cBits = RT_ALIGN_32(cBits, 32);
6959	__asm
6960	{
6961	# ifdef RT_ARCH_AMD64
6962	mov rdi, [pvBitmap]
6963	mov rbx, rdi
6964	# else
6965	mov edi, [pvBitmap]
6966	mov ebx, edi
6967	# endif
6968	mov edx, 0ffffffffh
6969	mov eax, edx
6970	mov ecx, [cBits]
6971	shr ecx, 5
6972	repe scasd
6973	je done
6974
6975	# ifdef RT_ARCH_AMD64
6976	lea rdi, [rdi - 4]
6977	xor eax, [rdi]
6978	sub rdi, rbx
6979	# else
6980	lea edi, [edi - 4]
6981	xor eax, [edi]
6982	sub edi, ebx
6983	# endif
6984	shl edi, 3
6985	bsf edx, eax
6986	add edx, edi
6987	done:
6988	mov [iBit], edx
6989	}
6990	# endif
6991	return iBit;
6992	}
6993	return -1;
6994	}
6995	#endif
6996
6997
6998	/**
6999	* Finds the next clear bit in a bitmap.
7000	*
7001	* @returns Index of the first zero bit.
7002	* @returns -1 if no clear bit was found.
7003	* @param pvBitmap Pointer to the bitmap (little endian).
7004	* @param cBits The number of bits in the bitmap. Multiple of 32.
7005	* @param iBitPrev The bit returned from the last search.
7006	* The search will start at iBitPrev + 1.
7007	*/
7008	#if RT_INLINE_ASM_EXTERNAL \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
7009	DECLASM(int) ASMBitNextClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_PROTO;
7010	#else
7011	DECLINLINE(int) ASMBitNextClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_DEF
7012	{
7013	const volatile uint32_t RT_FAR pau32Bitmap = (const volatile uint32_t RT_FAR )pvBitmap;
7014	int iBit = ++iBitPrev & 31;
7015	if (iBit)
7016	{
7017	/*
7018	* Inspect the 32-bit word containing the unaligned bit.
7019	*/
7020	uint32_t u32 = ~pau32Bitmap[iBitPrev / 32] >> iBit;
7021
7022	# if RT_INLINE_ASM_USES_INTRIN
7023	unsigned long ulBit = 0;
7024	if (_BitScanForward(&ulBit, u32))
7025	return ulBit + iBitPrev;
7026	# else
7027	# if RT_INLINE_ASM_GNU_STYLE
7028	__asm__ __volatile__("bsf %1, %0\n\t"
7029	"jnz 1f\n\t"
7030	"movl $-1, %0\n\t" /** @todo use conditional move for 64-bit? */
7031	"1:\n\t"
7032	: "=r" (iBit)
7033	: "r" (u32)
7034	: "cc");
7035	# else
7036	__asm
7037	{
7038	mov edx, [u32]
7039	bsf eax, edx
7040	jnz done
7041	mov eax, 0ffffffffh
7042	done:
7043	mov [iBit], eax
7044	}
7045	# endif
7046	if (iBit >= 0)
7047	return iBit + (int)iBitPrev;
7048	# endif
7049
7050	/*
7051	* Skip ahead and see if there is anything left to search.
7052	*/
7053	iBitPrev \|= 31;
7054	iBitPrev++;
7055	if (cBits <= (uint32_t)iBitPrev)
7056	return -1;
7057	}
7058
7059	/*
7060	* 32-bit aligned search, let ASMBitFirstClear do the dirty work.
7061	*/
7062	iBit = ASMBitFirstClear(&pau32Bitmap[iBitPrev / 32], cBits - iBitPrev);
7063	if (iBit >= 0)
7064	iBit += iBitPrev;
7065	return iBit;
7066	}
7067	#endif
7068
7069
7070	/**
7071	* Finds the first set bit in a bitmap.
7072	*
7073	* @returns Index of the first set bit.
7074	* @returns -1 if no clear bit was found.
7075	* @param pvBitmap Pointer to the bitmap (little endian).
7076	* @param cBits The number of bits in the bitmap. Multiple of 32.
7077	*/
7078	#if RT_INLINE_ASM_EXTERNAL \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
7079	DECLASM(int32_t) ASMBitFirstSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_PROTO;
7080	#else
7081	DECLINLINE(int32_t) ASMBitFirstSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_DEF
7082	{
7083	if (cBits)
7084	{
7085	int32_t iBit;
7086	# if RT_INLINE_ASM_GNU_STYLE
7087	RTCCUINTREG uEAX, uECX, uEDI;
7088	cBits = RT_ALIGN_32(cBits, 32);
7089	__asm__ __volatile__("repe; scasl\n\t"
7090	"je 1f\n\t"
7091	# ifdef RT_ARCH_AMD64
7092	"lea -4(%%rdi), %%rdi\n\t"
7093	"movl (%%rdi), %%eax\n\t"
7094	"subq %5, %%rdi\n\t"
7095	# else
7096	"lea -4(%%edi), %%edi\n\t"
7097	"movl (%%edi), %%eax\n\t"
7098	"subl %5, %%edi\n\t"
7099	# endif
7100	"shll $3, %%edi\n\t"
7101	"bsfl %%eax, %%edx\n\t"
7102	"addl %%edi, %%edx\n\t"
7103	"1:\t\n"
7104	: "=d" (iBit)
7105	, "=&c" (uECX)
7106	, "=&D" (uEDI)
7107	, "=&a" (uEAX)
7108	: "0" (0xffffffff)
7109	, "mr" (pvBitmap)
7110	, "1" (cBits >> 5)
7111	, "2" (pvBitmap)
7112	, "3" (0)
7113	: "cc");
7114	# else
7115	cBits = RT_ALIGN_32(cBits, 32);
7116	__asm
7117	{
7118	# ifdef RT_ARCH_AMD64
7119	mov rdi, [pvBitmap]
7120	mov rbx, rdi
7121	# else
7122	mov edi, [pvBitmap]
7123	mov ebx, edi
7124	# endif
7125	mov edx, 0ffffffffh
7126	xor eax, eax
7127	mov ecx, [cBits]
7128	shr ecx, 5
7129	repe scasd
7130	je done
7131	# ifdef RT_ARCH_AMD64
7132	lea rdi, [rdi - 4]
7133	mov eax, [rdi]
7134	sub rdi, rbx
7135	# else
7136	lea edi, [edi - 4]
7137	mov eax, [edi]
7138	sub edi, ebx
7139	# endif
7140	shl edi, 3
7141	bsf edx, eax
7142	add edx, edi
7143	done:
7144	mov [iBit], edx
7145	}
7146	# endif
7147	return iBit;
7148	}
7149	return -1;
7150	}
7151	#endif
7152
7153
7154	/**
7155	* Finds the next set bit in a bitmap.
7156	*
7157	* @returns Index of the next set bit.
7158	* @returns -1 if no set bit was found.
7159	* @param pvBitmap Pointer to the bitmap (little endian).
7160	* @param cBits The number of bits in the bitmap. Multiple of 32.
7161	* @param iBitPrev The bit returned from the last search.
7162	* The search will start at iBitPrev + 1.
7163	*/
7164	#if RT_INLINE_ASM_EXTERNAL \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
7165	DECLASM(int) ASMBitNextSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_PROTO;
7166	#else
7167	DECLINLINE(int) ASMBitNextSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_DEF
7168	{
7169	const volatile uint32_t RT_FAR pau32Bitmap = (const volatile uint32_t RT_FAR )pvBitmap;
7170	int iBit = ++iBitPrev & 31;
7171	if (iBit)
7172	{
7173	/*
7174	* Inspect the 32-bit word containing the unaligned bit.
7175	*/
7176	uint32_t u32 = pau32Bitmap[iBitPrev / 32] >> iBit;
7177
7178	# if RT_INLINE_ASM_USES_INTRIN
7179	unsigned long ulBit = 0;
7180	if (_BitScanForward(&ulBit, u32))
7181	return ulBit + iBitPrev;
7182	# else
7183	# if RT_INLINE_ASM_GNU_STYLE
7184	__asm__ __volatile__("bsf %1, %0\n\t"
7185	"jnz 1f\n\t" /** @todo use conditional move for 64-bit? */
7186	"movl $-1, %0\n\t"
7187	"1:\n\t"
7188	: "=r" (iBit)
7189	: "r" (u32)
7190	: "cc");
7191	# else
7192	__asm
7193	{
7194	mov edx, [u32]
7195	bsf eax, edx
7196	jnz done
7197	mov eax, 0ffffffffh
7198	done:
7199	mov [iBit], eax
7200	}
7201	# endif
7202	if (iBit >= 0)
7203	return iBit + (int)iBitPrev;
7204	# endif
7205
7206	/*
7207	* Skip ahead and see if there is anything left to search.
7208	*/
7209	iBitPrev \|= 31;
7210	iBitPrev++;
7211	if (cBits <= (uint32_t)iBitPrev)
7212	return -1;
7213	}
7214
7215	/*
7216	* 32-bit aligned search, let ASMBitFirstClear do the dirty work.
7217	*/
7218	iBit = ASMBitFirstSet(&pau32Bitmap[iBitPrev / 32], cBits - iBitPrev);
7219	if (iBit >= 0)
7220	iBit += iBitPrev;
7221	return iBit;
7222	}
7223	#endif
7224
7225
7226	/**
7227	* Finds the first bit which is set in the given 32-bit integer.
7228	* Bits are numbered from 1 (least significant) to 32.
7229	*
7230	* @returns index [1..32] of the first set bit.
7231	* @returns 0 if all bits are cleared.
7232	* @param u32 Integer to search for set bits.
7233	* @remarks Similar to ffs() in BSD.
7234	*/
7235	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7236	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitFirstSetU32(uint32_t u32) RT_NOTHROW_PROTO;
7237	#else
7238	DECLINLINE(unsigned) ASMBitFirstSetU32(uint32_t u32) RT_NOTHROW_DEF
7239	{
7240	# if RT_INLINE_ASM_USES_INTRIN
7241	unsigned long iBit;
7242	if (_BitScanForward(&iBit, u32))
7243	iBit++;
7244	else
7245	iBit = 0;
7246
7247	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7248	# if RT_INLINE_ASM_GNU_STYLE
7249	uint32_t iBit;
7250	__asm__ __volatile__("bsf %1, %0\n\t"
7251	"jnz 1f\n\t"
7252	"xorl %0, %0\n\t"
7253	"jmp 2f\n"
7254	"1:\n\t"
7255	"incl %0\n"
7256	"2:\n\t"
7257	: "=r" (iBit)
7258	: "rm" (u32)
7259	: "cc");
7260	# else
7261	uint32_t iBit;
7262	_asm
7263	{
7264	bsf eax, [u32]
7265	jnz found
7266	xor eax, eax
7267	jmp done
7268	found:
7269	inc eax
7270	done:
7271	mov [iBit], eax
7272	}
7273	# endif
7274
7275	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
7276	/*
7277	* Using the "count leading zeros (clz)" instruction here because there
7278	* is no dedicated instruction to get the first set bit.
7279	* Need to reverse the bits in the value with "rbit" first because
7280	* "clz" starts counting from the most significant bit.
7281	*/
7282	uint32_t iBit;
7283	__asm__ __volatile__(
7284	# if defined(RT_ARCH_ARM64)
7285	"rbit %w[uVal], %w[uVal]\n\t"
7286	"clz %w[iBit], %w[uVal]\n\t"
7287	# else
7288	"rbit %[uVal], %[uVal]\n\t"
7289	"clz %[iBit], %[uVal]\n\t"
7290	# endif
7291	: [uVal] "=r" (u32)
7292	, [iBit] "=r" (iBit)
7293	: "[uVal]" (u32));
7294	if (iBit != 32)
7295	iBit++;
7296	else
7297	iBit = 0; /* No bit set. */
7298
7299	# else
7300	# error "Port me"
7301	# endif
7302	return iBit;
7303	}
7304	#endif
7305
7306
7307	/**
7308	* Finds the first bit which is set in the given 32-bit integer.
7309	* Bits are numbered from 1 (least significant) to 32.
7310	*
7311	* @returns index [1..32] of the first set bit.
7312	* @returns 0 if all bits are cleared.
7313	* @param i32 Integer to search for set bits.
7314	* @remark Similar to ffs() in BSD.
7315	*/
7316	DECLINLINE(unsigned) ASMBitFirstSetS32(int32_t i32) RT_NOTHROW_DEF
7317	{
7318	return ASMBitFirstSetU32((uint32_t)i32);
7319	}
7320
7321
7322	/**
7323	* Finds the first bit which is set in the given 64-bit integer.
7324	*
7325	* Bits are numbered from 1 (least significant) to 64.
7326	*
7327	* @returns index [1..64] of the first set bit.
7328	* @returns 0 if all bits are cleared.
7329	* @param u64 Integer to search for set bits.
7330	* @remarks Similar to ffs() in BSD.
7331	*/
7332	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7333	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitFirstSetU64(uint64_t u64) RT_NOTHROW_PROTO;
7334	#else
7335	DECLINLINE(unsigned) ASMBitFirstSetU64(uint64_t u64) RT_NOTHROW_DEF
7336	{
7337	# if RT_INLINE_ASM_USES_INTRIN
7338	unsigned long iBit;
7339	# if ARCH_BITS == 64
7340	if (_BitScanForward64(&iBit, u64))
7341	iBit++;
7342	else
7343	iBit = 0;
7344	# else
7345	if (_BitScanForward(&iBit, (uint32_t)u64))
7346	iBit++;
7347	else if (_BitScanForward(&iBit, (uint32_t)(u64 >> 32)))
7348	iBit += 33;
7349	else
7350	iBit = 0;
7351	# endif
7352
7353	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
7354	uint64_t iBit;
7355	__asm__ __volatile__("bsfq %1, %0\n\t"
7356	"jnz 1f\n\t"
7357	"xorl %k0, %k0\n\t"
7358	"jmp 2f\n"
7359	"1:\n\t"
7360	"incl %k0\n"
7361	"2:\n\t"
7362	: "=r" (iBit)
7363	: "rm" (u64)
7364	: "cc");
7365
7366	# elif defined(RT_ARCH_ARM64)
7367	uint64_t iBit;
7368	__asm__ __volatile__("rbit %[uVal], %[uVal]\n\t"
7369	"clz %[iBit], %[uVal]\n\t"
7370	: [uVal] "=r" (u64)
7371	, [iBit] "=r" (iBit)
7372	: "[uVal]" (u64));
7373	if (iBit != 64)
7374	iBit++;
7375	else
7376	iBit = 0; /* No bit set. */
7377
7378	# else
7379	unsigned iBit = ASMBitFirstSetU32((uint32_t)u64);
7380	if (!iBit)
7381	{
7382	iBit = ASMBitFirstSetU32((uint32_t)(u64 >> 32));
7383	if (iBit)
7384	iBit += 32;
7385	}
7386	# endif
7387	return (unsigned)iBit;
7388	}
7389	#endif
7390
7391
7392	/**
7393	* Finds the first bit which is set in the given 16-bit integer.
7394	*
7395	* Bits are numbered from 1 (least significant) to 16.
7396	*
7397	* @returns index [1..16] of the first set bit.
7398	* @returns 0 if all bits are cleared.
7399	* @param u16 Integer to search for set bits.
7400	* @remarks For 16-bit bs3kit code.
7401	*/
7402	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7403	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitFirstSetU16(uint16_t u16) RT_NOTHROW_PROTO;
7404	#else
7405	DECLINLINE(unsigned) ASMBitFirstSetU16(uint16_t u16) RT_NOTHROW_DEF
7406	{
7407	return ASMBitFirstSetU32((uint32_t)u16);
7408	}
7409	#endif
7410
7411
7412	/**
7413	* Finds the last bit which is set in the given 32-bit integer.
7414	* Bits are numbered from 1 (least significant) to 32.
7415	*
7416	* @returns index [1..32] of the last set bit.
7417	* @returns 0 if all bits are cleared.
7418	* @param u32 Integer to search for set bits.
7419	* @remark Similar to fls() in BSD.
7420	*/
7421	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7422	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitLastSetU32(uint32_t u32) RT_NOTHROW_PROTO;
7423	#else
7424	DECLINLINE(unsigned) ASMBitLastSetU32(uint32_t u32) RT_NOTHROW_DEF
7425	{
7426	# if RT_INLINE_ASM_USES_INTRIN
7427	unsigned long iBit;
7428	if (_BitScanReverse(&iBit, u32))
7429	iBit++;
7430	else
7431	iBit = 0;
7432
7433	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7434	# if RT_INLINE_ASM_GNU_STYLE
7435	uint32_t iBit;
7436	__asm__ __volatile__("bsrl %1, %0\n\t"
7437	"jnz 1f\n\t"
7438	"xorl %0, %0\n\t"
7439	"jmp 2f\n"
7440	"1:\n\t"
7441	"incl %0\n"
7442	"2:\n\t"
7443	: "=r" (iBit)
7444	: "rm" (u32)
7445	: "cc");
7446	# else
7447	uint32_t iBit;
7448	_asm
7449	{
7450	bsr eax, [u32]
7451	jnz found
7452	xor eax, eax
7453	jmp done
7454	found:
7455	inc eax
7456	done:
7457	mov [iBit], eax
7458	}
7459	# endif
7460
7461	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
7462	uint32_t iBit;
7463	__asm__ __volatile__(
7464	# if defined(RT_ARCH_ARM64)
7465	"clz %w[iBit], %w[uVal]\n\t"
7466	# else
7467	"clz %[iBit], %[uVal]\n\t"
7468	# endif
7469	: [iBit] "=r" (iBit)
7470	: [uVal] "r" (u32));
7471	iBit = 32 - iBit;
7472
7473	# else
7474	# error "Port me"
7475	# endif
7476	return iBit;
7477	}
7478	#endif
7479
7480
7481	/**
7482	* Finds the last bit which is set in the given 32-bit integer.
7483	* Bits are numbered from 1 (least significant) to 32.
7484	*
7485	* @returns index [1..32] of the last set bit.
7486	* @returns 0 if all bits are cleared.
7487	* @param i32 Integer to search for set bits.
7488	* @remark Similar to fls() in BSD.
7489	*/
7490	DECLINLINE(unsigned) ASMBitLastSetS32(int32_t i32) RT_NOTHROW_DEF
7491	{
7492	return ASMBitLastSetU32((uint32_t)i32);
7493	}
7494
7495
7496	/**
7497	* Finds the last bit which is set in the given 64-bit integer.
7498	*
7499	* Bits are numbered from 1 (least significant) to 64.
7500	*
7501	* @returns index [1..64] of the last set bit.
7502	* @returns 0 if all bits are cleared.
7503	* @param u64 Integer to search for set bits.
7504	* @remark Similar to fls() in BSD.
7505	*/
7506	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7507	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitLastSetU64(uint64_t u64) RT_NOTHROW_PROTO;
7508	#else
7509	DECLINLINE(unsigned) ASMBitLastSetU64(uint64_t u64) RT_NOTHROW_DEF
7510	{
7511	# if RT_INLINE_ASM_USES_INTRIN
7512	unsigned long iBit;
7513	# if ARCH_BITS == 64
7514	if (_BitScanReverse64(&iBit, u64))
7515	iBit++;
7516	else
7517	iBit = 0;
7518	# else
7519	if (_BitScanReverse(&iBit, (uint32_t)(u64 >> 32)))
7520	iBit += 33;
7521	else if (_BitScanReverse(&iBit, (uint32_t)u64))
7522	iBit++;
7523	else
7524	iBit = 0;
7525	# endif
7526
7527	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
7528	uint64_t iBit;
7529	__asm__ __volatile__("bsrq %1, %0\n\t"
7530	"jnz 1f\n\t"
7531	"xorl %k0, %k0\n\t"
7532	"jmp 2f\n"
7533	"1:\n\t"
7534	"incl %k0\n"
7535	"2:\n\t"
7536	: "=r" (iBit)
7537	: "rm" (u64)
7538	: "cc");
7539
7540	# elif defined(RT_ARCH_ARM64)
7541	uint64_t iBit;
7542	__asm__ __volatile__("clz %[iBit], %[uVal]\n\t"
7543	: [iBit] "=r" (iBit)
7544	: [uVal] "r" (u64));
7545	iBit = 64 - iBit;
7546
7547	# else
7548	unsigned iBit = ASMBitLastSetU32((uint32_t)(u64 >> 32));
7549	if (iBit)
7550	iBit += 32;
7551	else
7552	iBit = ASMBitLastSetU32((uint32_t)u64);
7553	# endif
7554	return (unsigned)iBit;
7555	}
7556	#endif
7557
7558
7559	/**
7560	* Finds the last bit which is set in the given 16-bit integer.
7561	*
7562	* Bits are numbered from 1 (least significant) to 16.
7563	*
7564	* @returns index [1..16] of the last set bit.
7565	* @returns 0 if all bits are cleared.
7566	* @param u16 Integer to search for set bits.
7567	* @remarks For 16-bit bs3kit code.
7568	*/
7569	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7570	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitLastSetU16(uint16_t u16) RT_NOTHROW_PROTO;
7571	#else
7572	DECLINLINE(unsigned) ASMBitLastSetU16(uint16_t u16) RT_NOTHROW_DEF
7573	{
7574	return ASMBitLastSetU32((uint32_t)u16);
7575	}
7576	#endif
7577
7578
7579	/**
7580	* Count the number of leading zero bits in the given 32-bit integer.
7581	*
7582	* The counting starts with the most significate bit.
7583	*
7584	* @returns Number of most significant zero bits.
7585	* @returns 32 if all bits are cleared.
7586	* @param u32 Integer to consider.
7587	* @remarks Similar to __builtin_clz() in gcc, except defined zero input result.
7588	*/
7589	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7590	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountLeadingZerosU32(uint32_t u32) RT_NOTHROW_PROTO;
7591	#else
7592	DECLINLINE(unsigned) ASMCountLeadingZerosU32(uint32_t u32) RT_NOTHROW_DEF
7593	{
7594	# if RT_INLINE_ASM_USES_INTRIN
7595	unsigned long iBit;
7596	if (!_BitScanReverse(&iBit, u32))
7597	return 32;
7598	return 31 - (unsigned)iBit;
7599
7600	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7601	uint32_t iBit;
7602	# if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) && 0 /* significantly slower on 10980xe; 929 vs 237 ps/call */
7603	__asm__ __volatile__("bsrl %1, %0\n\t"
7604	"cmovzl %2, %0\n\t"
7605	: "=&r" (iBit)
7606	: "rm" (u32)
7607	, "rm" ((int32_t)-1)
7608	: "cc");
7609	# elif RT_INLINE_ASM_GNU_STYLE
7610	__asm__ __volatile__("bsr %1, %0\n\t"
7611	"jnz 1f\n\t"
7612	"mov $-1, %0\n\t"
7613	"1:\n\t"
7614	: "=r" (iBit)
7615	: "rm" (u32)
7616	: "cc");
7617	# else
7618	_asm
7619	{
7620	bsr eax, [u32]
7621	jnz found
7622	mov eax, -1
7623	found:
7624	mov [iBit], eax
7625	}
7626	# endif
7627	return 31 - iBit;
7628
7629	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
7630	uint32_t iBit;
7631	__asm__ __volatile__(
7632	# if defined(RT_ARCH_ARM64)
7633	"clz %w[iBit], %w[uVal]\n\t"
7634	# else
7635	"clz %[iBit], %[uVal]\n\t"
7636	# endif
7637	: [uVal] "=r" (u32)
7638	, [iBit] "=r" (iBit)
7639	: "[uVal]" (u32));
7640	return iBit;
7641
7642	# elif defined(__GNUC__)
7643	AssertCompile(sizeof(u32) == sizeof(unsigned int));
7644	return u32 ? __builtin_clz(u32) : 32;
7645
7646	# else
7647	# error "Port me"
7648	# endif
7649	}
7650	#endif
7651
7652
7653	/**
7654	* Count the number of leading zero bits in the given 64-bit integer.
7655	*
7656	* The counting starts with the most significate bit.
7657	*
7658	* @returns Number of most significant zero bits.
7659	* @returns 64 if all bits are cleared.
7660	* @param u64 Integer to consider.
7661	* @remarks Similar to __builtin_clzl() in gcc, except defined zero input
7662	* result.
7663	*/
7664	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7665	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountLeadingZerosU64(uint64_t u64) RT_NOTHROW_PROTO;
7666	#else
7667	DECLINLINE(unsigned) ASMCountLeadingZerosU64(uint64_t u64) RT_NOTHROW_DEF
7668	{
7669	# if RT_INLINE_ASM_USES_INTRIN
7670	unsigned long iBit;
7671	# if ARCH_BITS == 64
7672	if (_BitScanReverse64(&iBit, u64))
7673	return 63 - (unsigned)iBit;
7674	# else
7675	if (_BitScanReverse(&iBit, (uint32_t)(u64 >> 32)))
7676	return 31 - (unsigned)iBit;
7677	if (_BitScanReverse(&iBit, (uint32_t)u64))
7678	return 63 - (unsigned)iBit;
7679	# endif
7680	return 64;
7681
7682	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
7683	uint64_t iBit;
7684	# if 0 /* 10980xe benchmark: 932 ps/call - the slower variant */
7685	__asm__ __volatile__("bsrq %1, %0\n\t"
7686	"cmovzq %2, %0\n\t"
7687	: "=&r" (iBit)
7688	: "rm" (u64)
7689	, "rm" ((int64_t)-1)
7690	: "cc");
7691	# else /* 10980xe benchmark: 262 ps/call */
7692	__asm__ __volatile__("bsrq %1, %0\n\t"
7693	"jnz 1f\n\t"
7694	"mov $-1, %0\n\t"
7695	"1:\n\t"
7696	: "=&r" (iBit)
7697	: "rm" (u64)
7698	: "cc");
7699	# endif
7700	return 63 - (unsigned)iBit;
7701
7702	# elif defined(RT_ARCH_ARM64)
7703	uint64_t iBit;
7704	__asm__ __volatile__("clz %[iBit], %[uVal]\n\t"
7705	: [uVal] "=r" (u64)
7706	, [iBit] "=r" (iBit)
7707	: "[uVal]" (u64));
7708	return (unsigned)iBit;
7709
7710	# elif defined(__GNUC__) && ARCH_BITS == 64
7711	AssertCompile(sizeof(u64) == sizeof(unsigned long));
7712	return u64 ? __builtin_clzl(u64) : 64;
7713
7714	# else
7715	unsigned iBit = ASMCountLeadingZerosU32((uint32_t)(u64 >> 32));
7716	if (iBit == 32)
7717	iBit = ASMCountLeadingZerosU32((uint32_t)u64) + 32;
7718	return iBit;
7719	# endif
7720	}
7721	#endif
7722
7723
7724	/**
7725	* Count the number of leading zero bits in the given 16-bit integer.
7726	*
7727	* The counting starts with the most significate bit.
7728	*
7729	* @returns Number of most significant zero bits.
7730	* @returns 16 if all bits are cleared.
7731	* @param u16 Integer to consider.
7732	*/
7733	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7734	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountLeadingZerosU16(uint16_t u16) RT_NOTHROW_PROTO;
7735	#else
7736	DECLINLINE(unsigned) ASMCountLeadingZerosU16(uint16_t u16) RT_NOTHROW_DEF
7737	{
7738	# if RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)) && 0 /* slower (10980xe: 987 vs 292 ps/call) */
7739	uint16_t iBit;
7740	__asm__ __volatile__("bsrw %1, %0\n\t"
7741	"jnz 1f\n\t"
7742	"mov $-1, %0\n\t"
7743	"1:\n\t"
7744	: "=r" (iBit)
7745	: "rm" (u16)
7746	: "cc");
7747	return 15 - (int16_t)iBit;
7748	# else
7749	return ASMCountLeadingZerosU32((uint32_t)u16) - 16;
7750	# endif
7751	}
7752	#endif
7753
7754
7755	/**
7756	* Count the number of trailing zero bits in the given 32-bit integer.
7757	*
7758	* The counting starts with the least significate bit, i.e. the zero bit.
7759	*
7760	* @returns Number of lest significant zero bits.
7761	* @returns 32 if all bits are cleared.
7762	* @param u32 Integer to consider.
7763	* @remarks Similar to __builtin_ctz() in gcc, except defined zero input result.
7764	*/
7765	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7766	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountTrailingZerosU32(uint32_t u32) RT_NOTHROW_PROTO;
7767	#else
7768	DECLINLINE(unsigned) ASMCountTrailingZerosU32(uint32_t u32) RT_NOTHROW_DEF
7769	{
7770	# if RT_INLINE_ASM_USES_INTRIN
7771	unsigned long iBit;
7772	if (!_BitScanForward(&iBit, u32))
7773	return 32;
7774	return (unsigned)iBit;
7775
7776	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7777	uint32_t iBit;
7778	# if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) && 0 /* significantly slower on 10980xe; 932 vs 240 ps/call */
7779	__asm__ __volatile__("bsfl %1, %0\n\t"
7780	"cmovzl %2, %0\n\t"
7781	: "=&r" (iBit)
7782	: "rm" (u32)
7783	, "rm" ((int32_t)32)
7784	: "cc");
7785	# elif RT_INLINE_ASM_GNU_STYLE
7786	__asm__ __volatile__("bsfl %1, %0\n\t"
7787	"jnz 1f\n\t"
7788	"mov $32, %0\n\t"
7789	"1:\n\t"
7790	: "=r" (iBit)
7791	: "rm" (u32)
7792	: "cc");
7793	# else
7794	_asm
7795	{
7796	bsf eax, [u32]
7797	jnz found
7798	mov eax, 32
7799	found:
7800	mov [iBit], eax
7801	}
7802	# endif
7803	return iBit;
7804
7805	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
7806	/* Invert the bits and use clz. */
7807	uint32_t iBit;
7808	__asm__ __volatile__(
7809	# if defined(RT_ARCH_ARM64)
7810	"rbit %w[uVal], %w[uVal]\n\t"
7811	"clz %w[iBit], %w[uVal]\n\t"
7812	# else
7813	"rbit %[uVal], %[uVal]\n\t"
7814	"clz %[iBit], %[uVal]\n\t"
7815	# endif
7816	: [uVal] "=r" (u32)
7817	, [iBit] "=r" (iBit)
7818	: "[uVal]" (u32));
7819	return iBit;
7820
7821	# elif defined(__GNUC__)
7822	AssertCompile(sizeof(u32) == sizeof(unsigned int));
7823	return u32 ? __builtin_ctz(u32) : 32;
7824
7825	# else
7826	# error "Port me"
7827	# endif
7828	}
7829	#endif
7830
7831
7832	/**
7833	* Count the number of trailing zero bits in the given 64-bit integer.
7834	*
7835	* The counting starts with the least significate bit.
7836	*
7837	* @returns Number of least significant zero bits.
7838	* @returns 64 if all bits are cleared.
7839	* @param u64 Integer to consider.
7840	* @remarks Similar to __builtin_ctzl() in gcc, except defined zero input
7841	* result.
7842	*/
7843	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7844	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountTrailingZerosU64(uint64_t u64) RT_NOTHROW_PROTO;
7845	#else
7846	DECLINLINE(unsigned) ASMCountTrailingZerosU64(uint64_t u64) RT_NOTHROW_DEF
7847	{
7848	# if RT_INLINE_ASM_USES_INTRIN
7849	unsigned long iBit;
7850	# if ARCH_BITS == 64
7851	if (_BitScanForward64(&iBit, u64))
7852	return (unsigned)iBit;
7853	# else
7854	if (_BitScanForward(&iBit, (uint32_t)u64))
7855	return (unsigned)iBit;
7856	if (_BitScanForward(&iBit, (uint32_t)(u64 >> 32)))
7857	return (unsigned)iBit + 32;
7858	# endif
7859	return 64;
7860
7861	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
7862	uint64_t iBit;
7863	# if 0 /* 10980xe benchmark: 932 ps/call - the slower variant */
7864	__asm__ __volatile__("bsfq %1, %0\n\t"
7865	"cmovzq %2, %0\n\t"
7866	: "=&r" (iBit)
7867	: "rm" (u64)
7868	, "rm" ((int64_t)64)
7869	: "cc");
7870	# else /* 10980xe benchmark: 262 ps/call */
7871	__asm__ __volatile__("bsfq %1, %0\n\t"
7872	"jnz 1f\n\t"
7873	"mov $64, %0\n\t"
7874	"1:\n\t"
7875	: "=&r" (iBit)
7876	: "rm" (u64)
7877	: "cc");
7878	# endif
7879	return (unsigned)iBit;
7880
7881	# elif defined(RT_ARCH_ARM64)
7882	/* Invert the bits and use clz. */
7883	uint64_t iBit;
7884	__asm__ __volatile__("rbit %[uVal], %[uVal]\n\t"
7885	"clz %[iBit], %[uVal]\n\t"
7886	: [uVal] "=r" (u64)
7887	, [iBit] "=r" (iBit)
7888	: "[uVal]" (u64));
7889	return (unsigned)iBit;
7890
7891	# elif defined(__GNUC__) && ARCH_BITS == 64
7892	AssertCompile(sizeof(u64) == sizeof(unsigned long));
7893	return u64 ? __builtin_ctzl(u64) : 64;
7894
7895	# else
7896	unsigned iBit = ASMCountTrailingZerosU32((uint32_t)u64);
7897	if (iBit == 32)
7898	iBit = ASMCountTrailingZerosU32((uint32_t)(u64 >> 32)) + 32;
7899	return iBit;
7900	# endif
7901	}
7902	#endif
7903
7904
7905	/**
7906	* Count the number of trailing zero bits in the given 16-bit integer.
7907	*
7908	* The counting starts with the most significate bit.
7909	*
7910	* @returns Number of most significant zero bits.
7911	* @returns 16 if all bits are cleared.
7912	* @param u16 Integer to consider.
7913	*/
7914	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7915	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountTrailingZerosU16(uint16_t u16) RT_NOTHROW_PROTO;
7916	#else
7917	DECLINLINE(unsigned) ASMCountTrailingZerosU16(uint16_t u16) RT_NOTHROW_DEF
7918	{
7919	# if RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)) && 0 /* slower (10980xe: 992 vs 349 ps/call) */
7920	uint16_t iBit;
7921	__asm__ __volatile__("bsfw %1, %0\n\t"
7922	"jnz 1f\n\t"
7923	"mov $16, %0\n\t"
7924	"1:\n\t"
7925	: "=r" (iBit)
7926	: "rm" (u16)
7927	: "cc");
7928	return iBit;
7929	# else
7930	return ASMCountTrailingZerosU32((uint32_t)u16 \| UINT32_C(0x10000));
7931	#endif
7932	}
7933	#endif
7934
7935
7936	/**
7937	* Rotate 32-bit unsigned value to the left by @a cShift.
7938	*
7939	* @returns Rotated value.
7940	* @param u32 The value to rotate.
7941	* @param cShift How many bits to rotate by.
7942	*/
7943	#ifdef __WATCOMC__
7944	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMRotateLeftU32(uint32_t u32, unsigned cShift) RT_NOTHROW_PROTO;
7945	#else
7946	DECLINLINE(uint32_t) ASMRotateLeftU32(uint32_t u32, uint32_t cShift) RT_NOTHROW_DEF
7947	{
7948	# if RT_INLINE_ASM_USES_INTRIN
7949	return _rotl(u32, cShift);
7950
7951	# elif RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
7952	__asm__ __volatile__("roll %b1, %0" : "=g" (u32) : "Ic" (cShift), "0" (u32) : "cc");
7953	return u32;
7954
7955	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
7956	__asm__ __volatile__(
7957	# if defined(RT_ARCH_ARM64)
7958	"ror %w[uRet], %w[uVal], %w[cShift]\n\t"
7959	# else
7960	"ror %[uRet], %[uVal], %[cShift]\n\t"
7961	# endif
7962	: [uRet] "=r" (u32)
7963	: [uVal] "[uRet]" (u32)
7964	, [cShift] "r" (32 - (cShift & 31))); /** @todo there is an immediate form here */
7965	return u32;
7966
7967	# else
7968	cShift &= 31;
7969	return (u32 << cShift) \| (u32 >> (32 - cShift));
7970	# endif
7971	}
7972	#endif
7973
7974
7975	/**
7976	* Rotate 32-bit unsigned value to the right by @a cShift.
7977	*
7978	* @returns Rotated value.
7979	* @param u32 The value to rotate.
7980	* @param cShift How many bits to rotate by.
7981	*/
7982	#ifdef __WATCOMC__
7983	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMRotateRightU32(uint32_t u32, unsigned cShift) RT_NOTHROW_PROTO;
7984	#else
7985	DECLINLINE(uint32_t) ASMRotateRightU32(uint32_t u32, uint32_t cShift) RT_NOTHROW_DEF
7986	{
7987	# if RT_INLINE_ASM_USES_INTRIN
7988	return _rotr(u32, cShift);
7989
7990	# elif RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
7991	__asm__ __volatile__("rorl %b1, %0" : "=g" (u32) : "Ic" (cShift), "0" (u32) : "cc");
7992	return u32;
7993
7994	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
7995	__asm__ __volatile__(
7996	# if defined(RT_ARCH_ARM64)
7997	"ror %w[uRet], %w[uVal], %w[cShift]\n\t"
7998	# else
7999	"ror %[uRet], %[uVal], %[cShift]\n\t"
8000	# endif
8001	: [uRet] "=r" (u32)
8002	: [uVal] "[uRet]" (u32)
8003	, [cShift] "r" (cShift & 31)); /** @todo there is an immediate form here */
8004	return u32;
8005
8006	# else
8007	cShift &= 31;
8008	return (u32 >> cShift) \| (u32 << (32 - cShift));
8009	# endif
8010	}
8011	#endif
8012
8013
8014	/**
8015	* Rotate 64-bit unsigned value to the left by @a cShift.
8016	*
8017	* @returns Rotated value.
8018	* @param u64 The value to rotate.
8019	* @param cShift How many bits to rotate by.
8020	*/
8021	DECLINLINE(uint64_t) ASMRotateLeftU64(uint64_t u64, uint32_t cShift) RT_NOTHROW_DEF
8022	{
8023	#if RT_INLINE_ASM_USES_INTRIN
8024	return _rotl64(u64, cShift);
8025
8026	#elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
8027	__asm__ __volatile__("rolq %b1, %0" : "=g" (u64) : "Jc" (cShift), "0" (u64) : "cc");
8028	return u64;
8029
8030	#elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_X86)
8031	uint32_t uSpill;
8032	__asm__ __volatile__("testb $0x20, %%cl\n\t" /* if (cShift >= 0x20) { swap(u64.hi, u64lo); cShift -= 0x20; } */
8033	"jz 1f\n\t"
8034	"xchgl %%eax, %%edx\n\t"
8035	"1:\n\t"
8036	"andb $0x1f, %%cl\n\t" /* if (cShift & 0x1f) { */
8037	"jz 2f\n\t"
8038	"movl %%edx, %2\n\t" /* save the hi value in %3. */
8039	"shldl %%cl,%%eax,%%edx\n\t" /* shift the hi value left, feeding MSBits from the low value. */
8040	"shldl %%cl,%2,%%eax\n\t" /* shift the lo value left, feeding MSBits from the saved hi value. */
8041	"2:\n\t" /* } */
8042	: "=A" (u64)
8043	, "=c" (cShift)
8044	, "=r" (uSpill)
8045	: "0" (u64)
8046	, "1" (cShift)
8047	: "cc");
8048	return u64;
8049
8050	# elif defined(RT_ARCH_ARM64)
8051	__asm__ __volatile__("ror %[uRet], %[uVal], %[cShift]\n\t"
8052	: [uRet] "=r" (u64)
8053	: [uVal] "[uRet]" (u64)
8054	, [cShift] "r" ((uint64_t)(64 - (cShift & 63)))); /** @todo there is an immediate form here */
8055	return u64;
8056
8057	#else
8058	cShift &= 63;
8059	return (u64 << cShift) \| (u64 >> (64 - cShift));
8060	#endif
8061	}
8062
8063
8064	/**
8065	* Rotate 64-bit unsigned value to the right by @a cShift.
8066	*
8067	* @returns Rotated value.
8068	* @param u64 The value to rotate.
8069	* @param cShift How many bits to rotate by.
8070	*/
8071	DECLINLINE(uint64_t) ASMRotateRightU64(uint64_t u64, uint32_t cShift) RT_NOTHROW_DEF
8072	{
8073	#if RT_INLINE_ASM_USES_INTRIN
8074	return _rotr64(u64, cShift);
8075
8076	#elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
8077	__asm__ __volatile__("rorq %b1, %0" : "=g" (u64) : "Jc" (cShift), "0" (u64) : "cc");
8078	return u64;
8079
8080	#elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_X86)
8081	uint32_t uSpill;
8082	__asm__ __volatile__("testb $0x20, %%cl\n\t" /* if (cShift >= 0x20) { swap(u64.hi, u64lo); cShift -= 0x20; } */
8083	"jz 1f\n\t"
8084	"xchgl %%eax, %%edx\n\t"
8085	"1:\n\t"
8086	"andb $0x1f, %%cl\n\t" /* if (cShift & 0x1f) { */
8087	"jz 2f\n\t"
8088	"movl %%edx, %2\n\t" /* save the hi value in %3. */
8089	"shrdl %%cl,%%eax,%%edx\n\t" /* shift the hi value right, feeding LSBits from the low value. */
8090	"shrdl %%cl,%2,%%eax\n\t" /* shift the lo value right, feeding LSBits from the saved hi value. */
8091	"2:\n\t" /* } */
8092	: "=A" (u64)
8093	, "=c" (cShift)
8094	, "=r" (uSpill)
8095	: "0" (u64)
8096	, "1" (cShift)
8097	: "cc");
8098	return u64;
8099
8100	# elif defined(RT_ARCH_ARM64)
8101	__asm__ __volatile__("ror %[uRet], %[uVal], %[cShift]\n\t"
8102	: [uRet] "=r" (u64)
8103	: [uVal] "[uRet]" (u64)
8104	, [cShift] "r" ((uint64_t)(cShift & 63))); /** @todo there is an immediate form here */
8105	return u64;
8106
8107	#else
8108	cShift &= 63;
8109	return (u64 >> cShift) \| (u64 << (64 - cShift));
8110	#endif
8111	}
8112
8113	/** @} */
8114
8115
8116	/** @} */
8117
8118	/*
8119	* Include #pragma aux definitions for Watcom C/C++.
8120	*/
8121	#if defined(__WATCOMC__) && ARCH_BITS == 16 && defined(RT_ARCH_X86)
8122	# define IPRT_ASM_WATCOM_X86_16_WITH_PRAGMAS
8123	# undef IPRT_INCLUDED_asm_watcom_x86_16_h
8124	# include "asm-watcom-x86-16.h"
8125	#elif defined(__WATCOMC__) && ARCH_BITS == 32 && defined(RT_ARCH_X86)
8126	# define IPRT_ASM_WATCOM_X86_32_WITH_PRAGMAS
8127	# undef IPRT_INCLUDED_asm_watcom_x86_32_h
8128	# include "asm-watcom-x86-32.h"
8129	#endif
8130
8131	#endif /* !IPRT_INCLUDED_asm_h */
8132

Note: See TracBrowser for help on using the repository browser.

source: vbox/trunk/include/iprt/asm.h@ 97778

Download in other formats: