asm.h@ 105688

Last change on this file since 105688 was 104807, checked in by vboxsync, 6 months ago
iprt/asm.h: re-shuffle declarations and code to meet ISO C90 expectations to avoid build failures in some contexts. bugref:10687
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 288.1 KB

Line
1	/** @file
2	* IPRT - Assembly Functions.
3	*/
4
5	/*
6	* Copyright (C) 2006-2023 Oracle and/or its affiliates.
7	*
8	* This file is part of VirtualBox base platform packages, as
9	* available from https://www.virtualbox.org.
10	*
11	* This program is free software; you can redistribute it and/or
12	* modify it under the terms of the GNU General Public License
13	* as published by the Free Software Foundation, in version 3 of the
14	* License.
15	*
16	* This program is distributed in the hope that it will be useful, but
17	* WITHOUT ANY WARRANTY; without even the implied warranty of
18	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19	* General Public License for more details.
20	*
21	* You should have received a copy of the GNU General Public License
22	* along with this program; if not, see <https://www.gnu.org/licenses>.
23	*
24	* The contents of this file may alternatively be used under the terms
25	* of the Common Development and Distribution License Version 1.0
26	* (CDDL), a copy of it is provided in the "COPYING.CDDL" file included
27	* in the VirtualBox distribution, in which case the provisions of the
28	* CDDL are applicable instead of those of the GPL.
29	*
30	* You may elect to license modified versions of this file under the
31	* terms and conditions of either the GPL or the CDDL or both.
32	*
33	* SPDX-License-Identifier: GPL-3.0-only OR CDDL-1.0
34	*/
35
36	#ifndef IPRT_INCLUDED_asm_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	#if (defined(RT_ARCH_ARM64) && defined(RT_OS_DARWIN)) \|\| defined(DOXYGEN_RUNNING)
102	/** @def RTASM_ARM64_USE_FEAT_LSE
103	* Use instructions from the FEAT_LSE set to implement atomic operations,
104	* assuming that the host CPU always supports these. */
105	# define RTASM_ARM64_USE_FEAT_LSE 1
106	/** @def RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB
107	* Set to use DMB w/o barrier in most places and rely on the acquire-release
108	* aspects to do the serializing. The assumption is that the tstRTInline
109	* benchmark may be skewing the results testing an unusual scenario. */
110	# define RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB 1
111	#endif
112
113
114	/*
115	* Undefine all symbols we have Watcom C/C++ #pragma aux'es for.
116	*/
117	#if defined(__WATCOMC__) && ARCH_BITS == 16 && defined(RT_ARCH_X86)
118	# include "asm-watcom-x86-16.h"
119	#elif defined(__WATCOMC__) && ARCH_BITS == 32 && defined(RT_ARCH_X86)
120	# include "asm-watcom-x86-32.h"
121	#endif
122
123
124	/** @defgroup grp_rt_asm ASM - Assembly Routines
125	* @ingroup grp_rt
126	*
127	* @remarks The difference between ordered and unordered atomic operations are
128	* that the former will complete outstanding reads and writes before
129	* continuing while the latter doesn't make any promises about the
130	* order. Ordered operations doesn't, it seems, make any 100% promise
131	* wrt to whether the operation will complete before any subsequent
132	* memory access. (please, correct if wrong.)
133	*
134	* ASMAtomicSomething operations are all ordered, while
135	* ASMAtomicUoSomething are unordered (note the Uo).
136	*
137	* Please note that ordered operations does not necessarily imply a
138	* compiler (memory) barrier. The user has to use the
139	* ASMCompilerBarrier() macro when that is deemed necessary.
140	*
141	* @remarks Some remarks about __volatile__: Without this keyword gcc is allowed
142	* to reorder or even optimize assembler instructions away. For
143	* instance, in the following code the second rdmsr instruction is
144	* optimized away because gcc treats that instruction as deterministic:
145	*
146	* @code
147	* static inline uint64_t rdmsr_low(int idx)
148	* {
149	* uint32_t low;
150	* __asm__ ("rdmsr" : "=a"(low) : "c"(idx) : "edx");
151	* }
152	* ...
153	* uint32_t msr1 = rdmsr_low(1);
154	* foo(msr1);
155	* msr1 = rdmsr_low(1);
156	* bar(msr1);
157	* @endcode
158	*
159	* The input parameter of rdmsr_low is the same for both calls and
160	* therefore gcc will use the result of the first call as input
161	* parameter for bar() as well. For rdmsr this is not acceptable as
162	* this instruction is _not_ deterministic. This applies to reading
163	* machine status information in general.
164	*
165	* @{
166	*/
167
168
169	/** @def RT_INLINE_ASM_GCC_4_3_X_X86
170	* Used to work around some 4.3.x register allocation issues in this version of
171	* the compiler. So far this workaround is still required for 4.4 and 4.5 but
172	* definitely not for 5.x */
173	#if (RT_GNUC_PREREQ(4, 3) && !RT_GNUC_PREREQ(5, 0) && defined(__i386__))
174	# define RT_INLINE_ASM_GCC_4_3_X_X86 1
175	#else
176	# define RT_INLINE_ASM_GCC_4_3_X_X86 0
177	#endif
178
179	/** @def RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
180	* i686-apple-darwin9-gcc-4.0.1 (GCC) 4.0.1 (Apple Inc. build 5493) screws up
181	* RTSemRWRequestWrite semsemrw-lockless-generic.cpp in release builds. PIC
182	* mode, x86.
183	*
184	* Some gcc 4.3.x versions may have register allocation issues with cmpxchg8b
185	* when in PIC mode on x86.
186	*/
187	#ifndef RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
188	# if defined(DOXYGEN_RUNNING) \|\| defined(__WATCOMC__) /* Watcom has trouble with the expression below */
189	# define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 1
190	# elif defined(_MSC_VER) /* Visual C++ has trouble too, but it'll only tell us when C4688 is enabled. */
191	# define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 0
192	# elif ( (defined(PIC) \|\| defined(__PIC__)) \
193	&& defined(RT_ARCH_X86) \
194	&& ( RT_INLINE_ASM_GCC_4_3_X_X86 \
195	\|\| defined(RT_OS_DARWIN)) )
196	# define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 1
197	# else
198	# define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 0
199	# endif
200	#endif
201
202
203	/*
204	* ARM is great fun.
205	*/
206	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
207
208	# define RTASM_ARM_NO_BARRIER
209	# ifdef RT_ARCH_ARM64
210	# define RTASM_ARM_NO_BARRIER_IN_REG
211	# define RTASM_ARM_NO_BARRIER_COMMA_IN_REG
212	# define RTASM_ARM_DSB_SY "dsb sy\n\t"
213	# define RTASM_ARM_DSB_SY_IN_REG
214	# define RTASM_ARM_DSB_SY_COMMA_IN_REG
215	# define RTASM_ARM_DMB_SY "dmb sy\n\t"
216	# define RTASM_ARM_DMB_SY_IN_REG
217	# define RTASM_ARM_DMB_SY_COMMA_IN_REG
218	# define RTASM_ARM_DMB_ST "dmb st\n\t"
219	# define RTASM_ARM_DMB_ST_IN_REG
220	# define RTASM_ARM_DMB_ST_COMMA_IN_REG
221	# define RTASM_ARM_DMB_LD "dmb ld\n\t"
222	# define RTASM_ARM_DMB_LD_IN_REG
223	# define RTASM_ARM_DMB_LD_COMMA_IN_REG
224	# define RTASM_ARM_PICK_6432(expr64, expr32) expr64
225	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \
226	uint32_t rcSpill; \
227	uint32_t u32NewRet; \
228	__asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \
229	RTASM_ARM_##barrier_type /* before lable? */ \
230	"ldaxr %w[uNew], %[pMem]\n\t" \
231	modify64 \
232	"stlxr %w[rc], %w[uNew], %[pMem]\n\t" \
233	"cbnz %w[rc], Ltry_again_" #name "_%=\n\t" \
234	: [pMem] "+Q" (*a_pu32Mem) \
235	, [uNew] "=&r" (u32NewRet) \
236	, [rc] "=&r" (rcSpill) \
237	: in_reg \
238	: "cc")
239	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \
240	uint32_t rcSpill; \
241	uint32_t u32OldRet; \
242	uint32_t u32NewSpill; \
243	__asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \
244	RTASM_ARM_##barrier_type /* before lable? */ \
245	"ldaxr %w[uOld], %[pMem]\n\t" \
246	modify64 \
247	"stlxr %w[rc], %w[uNew], %[pMem]\n\t" \
248	"cbnz %w[rc], Ltry_again_" #name "_%=\n\t" \
249	: [pMem] "+Q" (*a_pu32Mem) \
250	, [uOld] "=&r" (u32OldRet) \
251	, [uNew] "=&r" (u32NewSpill) \
252	, [rc] "=&r" (rcSpill) \
253	: in_reg \
254	: "cc")
255	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \
256	uint32_t rcSpill; \
257	uint64_t u64NewRet; \
258	__asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \
259	RTASM_ARM_##barrier_type /* before lable? */ \
260	"ldaxr %[uNew], %[pMem]\n\t" \
261	modify64 \
262	"stlxr %w[rc], %[uNew], %[pMem]\n\t" \
263	"cbnz %w[rc], Ltry_again_" #name "_%=\n\t" \
264	: [pMem] "+Q" (*a_pu64Mem) \
265	, [uNew] "=&r" (u64NewRet) \
266	, [rc] "=&r" (rcSpill) \
267	: in_reg \
268	: "cc")
269	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \
270	uint32_t rcSpill; \
271	uint64_t u64OldRet; \
272	uint64_t u64NewSpill; \
273	__asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \
274	RTASM_ARM_##barrier_type /* before lable? */ \
275	"ldaxr %[uOld], %[pMem]\n\t" \
276	modify64 \
277	"stlxr %w[rc], %[uNew], %[pMem]\n\t" \
278	"cbnz %w[rc], Ltry_again_" #name "_%=\n\t" \
279	: [pMem] "+Q" (*a_pu64Mem) \
280	, [uOld] "=&r" (u64OldRet) \
281	, [uNew] "=&r" (u64NewSpill) \
282	, [rc] "=&r" (rcSpill) \
283	: in_reg \
284	: "cc")
285
286	# else /* RT_ARCH_ARM32 */
287	# define RTASM_ARM_PICK_6432(expr64, expr32) expr32
288	# if RT_ARCH_ARM32 >= 7
289	# warning armv7
290	# define RTASM_ARM_NO_BARRIER_IN_REG
291	# define RTASM_ARM_NO_BARRIER_COMMA_IN_REG
292	# define RTASM_ARM_DSB_SY "dsb sy\n\t"
293	# define RTASM_ARM_DSB_SY_IN_REG "X" (0xfade)
294	# define RTASM_ARM_DMB_SY "dmb sy\n\t"
295	# define RTASM_ARM_DMB_SY_IN_REG "X" (0xfade)
296	# define RTASM_ARM_DMB_ST "dmb st\n\t"
297	# define RTASM_ARM_DMB_ST_IN_REG "X" (0xfade)
298	# define RTASM_ARM_DMB_LD "dmb ld\n\t"
299	# define RTASM_ARM_DMB_LD_IN_REG "X" (0xfade)
300
301	# elif RT_ARCH_ARM32 >= 6
302	# warning armv6
303	# define RTASM_ARM_DSB_SY "mcr p15, 0, %[uZero], c7, c10, 4\n\t"
304	# define RTASM_ARM_DSB_SY_IN_REG [uZero] "r" (0)
305	# define RTASM_ARM_DMB_SY "mcr p15, 0, %[uZero], c7, c10, 5\n\t"
306	# define RTASM_ARM_DMB_SY_IN_REG [uZero] "r" (0)
307	# define RTASM_ARM_DMB_ST RTASM_ARM_DMB_SY
308	# define RTASM_ARM_DMB_ST_IN_REG RTASM_ARM_DMB_SY_IN_REG
309	# define RTASM_ARM_DMB_LD RTASM_ARM_DMB_SY
310	# define RTASM_ARM_DMB_LD_IN_REG RTASM_ARM_DMB_SY_IN_REG
311
312	# elif RT_ARCH_ARM32 >= 4
313	# warning armv5 or older
314	# define RTASM_ARM_DSB_SY "mcr p15, 0, %[uZero], c7, c10, 4\n\t"
315	# define RTASM_ARM_DSB_SY_IN_REG [uZero] "r" (0)
316	# define RTASM_ARM_DMB_SY RTASM_ARM_DSB_SY
317	# define RTASM_ARM_DMB_SY_IN_REG RTASM_ARM_DSB_SY_IN_REG
318	# define RTASM_ARM_DMB_ST RTASM_ARM_DSB_SY
319	# define RTASM_ARM_DMB_ST_IN_REG RTASM_ARM_DSB_SY_IN_REG
320	# define RTASM_ARM_DMB_LD RTASM_ARM_DSB_SY
321	# define RTASM_ARM_DMB_LD_IN_REG RTASM_ARM_DSB_SY_IN_REG
322	# else
323	# error "huh? Odd RT_ARCH_ARM32 value!"
324	# endif
325	# define RTASM_ARM_DSB_SY_COMMA_IN_REG , RTASM_ARM_DSB_SY_IN_REG
326	# define RTASM_ARM_DMB_SY_COMMA_IN_REG , RTASM_ARM_DMB_SY_IN_REG
327	# define RTASM_ARM_DMB_ST_COMMA_IN_REG , RTASM_ARM_DMB_ST_IN_REG
328	# define RTASM_ARM_DMB_LD_COMMA_IN_REG , RTASM_ARM_DMB_LD_IN_REG
329	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \
330	uint32_t rcSpill; \
331	uint32_t u32NewRet; \
332	__asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \
333	RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \
334	"ldrex %[uNew], %[pMem]\n\t" \
335	modify32 \
336	"strex %[rc], %[uNew], %[pMem]\n\t" \
337	"cmp %[rc], #0\n\t" \
338	"bne Ltry_again_" #name "_%=\n\t" \
339	: [pMem] "+m" (*a_pu32Mem) \
340	, [uNew] "=&r" (u32NewRet) \
341	, [rc] "=&r" (rcSpill) \
342	: RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \
343	, in_reg \
344	: "cc")
345	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \
346	uint32_t rcSpill; \
347	uint32_t u32OldRet; \
348	uint32_t u32NewSpill; \
349	__asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \
350	RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \
351	"ldrex %[uOld], %[pMem]\n\t" \
352	modify32 \
353	"strex %[rc], %[uNew], %[pMem]\n\t" \
354	"cmp %[rc], #0\n\t" \
355	"bne Ltry_again_" #name "_%=\n\t" \
356	: [pMem] "+m" (*a_pu32Mem) \
357	, [uOld] "=&r" (u32OldRet) \
358	, [uNew] "=&r" (u32NewSpill) \
359	, [rc] "=&r" (rcSpill) \
360	: RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \
361	, in_reg \
362	: "cc")
363	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \
364	uint32_t rcSpill; \
365	uint64_t u64NewRet; \
366	__asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \
367	RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \
368	"ldrexd %[uNew], %H[uNew], %[pMem]\n\t" \
369	modify32 \
370	"strexd %[rc], %[uNew], %H[uNew], %[pMem]\n\t" \
371	"cmp %[rc], #0\n\t" \
372	"bne Ltry_again_" #name "_%=\n\t" \
373	: [pMem] "+m" (*a_pu64Mem), \
374	[uNew] "=&r" (u64NewRet), \
375	[rc] "=&r" (rcSpill) \
376	: RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \
377	, in_reg \
378	: "cc")
379	# define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \
380	uint32_t rcSpill; \
381	uint64_t u64OldRet; \
382	uint64_t u64NewSpill; \
383	__asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \
384	RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \
385	"ldrexd %[uOld], %H[uOld], %[pMem]\n\t" \
386	modify32 \
387	"strexd %[rc], %[uNew], %H[uNew], %[pMem]\n\t" \
388	"cmp %[rc], #0\n\t" \
389	"bne Ltry_again_" #name "_%=\n\t" \
390	: [pMem] "+m" (*a_pu64Mem), \
391	[uOld] "=&r" (u64OldRet), \
392	[uNew] "=&r" (u64NewSpill), \
393	[rc] "=&r" (rcSpill) \
394	: RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \
395	, in_reg \
396	: "cc")
397	# endif /* RT_ARCH_ARM32 */
398	#endif
399
400
401	/** @def ASMReturnAddress
402	* Gets the return address of the current (or calling if you like) function or method.
403	*/
404	#ifdef _MSC_VER
405	# ifdef __cplusplus
406	extern "C"
407	# endif
408	void * _ReturnAddress(void);
409	# pragma intrinsic(_ReturnAddress)
410	# define ASMReturnAddress() _ReturnAddress()
411	#elif defined(__GNUC__) \|\| defined(DOXYGEN_RUNNING)
412	# define ASMReturnAddress() __builtin_return_address(0)
413	#elif defined(__WATCOMC__)
414	# define ASMReturnAddress() Watcom_does_not_appear_to_have_intrinsic_return_address_function()
415	#else
416	# error "Unsupported compiler."
417	#endif
418
419
420	/**
421	* Compiler memory barrier.
422	*
423	* Ensure that the compiler does not use any cached (register/tmp stack) memory
424	* values or any outstanding writes when returning from this function.
425	*
426	* This function must be used if non-volatile data is modified by a
427	* device or the VMM. Typical cases are port access, MMIO access,
428	* trapping instruction, etc.
429	*/
430	#if RT_INLINE_ASM_GNU_STYLE
431	# define ASMCompilerBarrier() do { __asm__ __volatile__("" : : : "memory"); } while (0)
432	#elif RT_INLINE_ASM_USES_INTRIN
433	# define ASMCompilerBarrier() do { _ReadWriteBarrier(); } while (0)
434	#elif defined(__WATCOMC__)
435	void ASMCompilerBarrier(void);
436	#else /* 2003 should have _ReadWriteBarrier() but I guess we're at 2002 level then... */
437	DECLINLINE(void) ASMCompilerBarrier(void) RT_NOTHROW_DEF
438	{
439	__asm
440	{
441	}
442	}
443	#endif
444
445
446	/** @def ASMBreakpoint
447	* Debugger Breakpoint.
448	* @deprecated Use RT_BREAKPOINT instead.
449	* @internal
450	*/
451	#define ASMBreakpoint() RT_BREAKPOINT()
452
453
454	/**
455	* Spinloop hint for platforms that have these, empty function on the other
456	* platforms.
457	*
458	* x86 & AMD64: The PAUSE variant of NOP for helping hyperthreaded CPUs detecting
459	* spin locks.
460	*/
461	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
462	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMNopPause(void) RT_NOTHROW_PROTO;
463	#else
464	DECLINLINE(void) ASMNopPause(void) RT_NOTHROW_DEF
465	{
466	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
467	# if RT_INLINE_ASM_GNU_STYLE
468	__asm__ __volatile__(".byte 0xf3,0x90\n\t");
469	# else
470	__asm {
471	_emit 0f3h
472	_emit 090h
473	}
474	# endif
475
476	# elif defined(RT_ARCH_ARM32) \|\| defined(RT_ARCH_ARM64)
477	__asm__ __volatile__("yield\n\t"); /* ARMv6K+ */
478
479	# else
480	/* dummy */
481	# endif
482	}
483	#endif
484
485
486	/**
487	* Atomically Exchange an unsigned 8-bit value, ordered.
488	*
489	* @returns Current *pu8 value
490	* @param pu8 Pointer to the 8-bit variable to update.
491	* @param u8 The 8-bit value to assign to *pu8.
492	*/
493	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
494	RT_ASM_DECL_PRAGMA_WATCOM(uint8_t) ASMAtomicXchgU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_PROTO;
495	#else
496	DECLINLINE(uint8_t) ASMAtomicXchgU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_DEF
497	{
498	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
499	# if RT_INLINE_ASM_GNU_STYLE
500	__asm__ __volatile__("xchgb %0, %1\n\t"
501	: "=m" (*pu8)
502	, "=q" (u8) /* =r - busted on g++ (GCC) 3.4.4 20050721 (Red Hat 3.4.4-2) */
503	: "1" (u8)
504	, "m" (*pu8));
505	# else
506	__asm
507	{
508	# ifdef RT_ARCH_AMD64
509	mov rdx, [pu8]
510	mov al, [u8]
511	xchg [rdx], al
512	mov [u8], al
513	# else
514	mov edx, [pu8]
515	mov al, [u8]
516	xchg [edx], al
517	mov [u8], al
518	# endif
519	}
520	# endif
521	return u8;
522
523	# elif defined(RT_ARCH_ARM32) \|\| defined(RT_ARCH_ARM64)
524	uint32_t uOld;
525	# if defined(RTASM_ARM64_USE_FEAT_LSE)
526	/* SWPALB is ~40% more expensive than the non-LSE variant (M1), but since we
527	have the barrier we shouldn't need that, right? Ordering should be taken
528	care of by the DMB. The SWPB is rather cheap (~70% faster). */
529	__asm__ __volatile__("Lstart_ASMAtomicXchgU8_%=:\n\t"
530	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
531	"swpalb %w[uNew], %w[uOld], %[pMem]\n\t"
532	# else
533	RTASM_ARM_DMB_SY
534	"swpb %w[uNew], %w[uOld], %[pMem]\n\t"
535	# endif
536	: [pMem] "+Q" (*pu8)
537	, [uOld] "=&r" (uOld)
538	: [uNew] "r" ((uint32_t)u8)
539	: );
540	# else
541	uint32_t rcSpill;
542	__asm__ __volatile__("Ltry_again_ASMAtomicXchgU8_%=:\n\t"
543	RTASM_ARM_DMB_SY
544	# if defined(RT_ARCH_ARM64)
545	"ldaxrb %w[uOld], %[pMem]\n\t"
546	"stlxrb %w[rc], %w[uNew], %[pMem]\n\t"
547	"cbnz %w[rc], Ltry_again_ASMAtomicXchgU8_%=\n\t"
548	# else
549	"ldrexb %[uOld], %[pMem]\n\t" /* ARMv6+ */
550	"strexb %[rc], %[uNew], %[pMem]\n\t"
551	"cmp %[rc], #0\n\t"
552	"bne Ltry_again_ASMAtomicXchgU8_%=\n\t"
553	# endif
554	: [pMem] "+Q" (*pu8)
555	, [uOld] "=&r" (uOld)
556	, [rc] "=&r" (rcSpill)
557	: [uNew] "r" ((uint32_t)u8)
558	RTASM_ARM_DMB_SY_COMMA_IN_REG
559	: "cc");
560	# endif
561	return (uint8_t)uOld;
562
563	# else
564	# error "Port me"
565	# endif
566	}
567	#endif
568
569
570	/**
571	* Atomically Exchange a signed 8-bit value, ordered.
572	*
573	* @returns Current *pu8 value
574	* @param pi8 Pointer to the 8-bit variable to update.
575	* @param i8 The 8-bit value to assign to *pi8.
576	*/
577	DECLINLINE(int8_t) ASMAtomicXchgS8(volatile int8_t RT_FAR *pi8, int8_t i8) RT_NOTHROW_DEF
578	{
579	return (int8_t)ASMAtomicXchgU8((volatile uint8_t RT_FAR *)pi8, (uint8_t)i8);
580	}
581
582
583	/**
584	* Atomically Exchange a bool value, ordered.
585	*
586	* @returns Current *pf value
587	* @param pf Pointer to the 8-bit variable to update.
588	* @param f The 8-bit value to assign to *pi8.
589	*/
590	DECLINLINE(bool) ASMAtomicXchgBool(volatile bool RT_FAR *pf, bool f) RT_NOTHROW_DEF
591	{
592	#ifdef _MSC_VER
593	return !!ASMAtomicXchgU8((volatile uint8_t RT_FAR *)pf, (uint8_t)f);
594	#else
595	return (bool)ASMAtomicXchgU8((volatile uint8_t RT_FAR *)pf, (uint8_t)f);
596	#endif
597	}
598
599
600	/**
601	* Atomically Exchange an unsigned 16-bit value, ordered.
602	*
603	* @returns Current *pu16 value
604	* @param pu16 Pointer to the 16-bit variable to update.
605	* @param u16 The 16-bit value to assign to *pu16.
606	*/
607	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
608	RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMAtomicXchgU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_PROTO;
609	#else
610	DECLINLINE(uint16_t) ASMAtomicXchgU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_DEF
611	{
612	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
613	# if RT_INLINE_ASM_GNU_STYLE
614	__asm__ __volatile__("xchgw %0, %1\n\t"
615	: "=m" (*pu16)
616	, "=r" (u16)
617	: "1" (u16)
618	, "m" (*pu16));
619	# else
620	__asm
621	{
622	# ifdef RT_ARCH_AMD64
623	mov rdx, [pu16]
624	mov ax, [u16]
625	xchg [rdx], ax
626	mov [u16], ax
627	# else
628	mov edx, [pu16]
629	mov ax, [u16]
630	xchg [edx], ax
631	mov [u16], ax
632	# endif
633	}
634	# endif
635	return u16;
636
637	# elif defined(RT_ARCH_ARM32) \|\| defined(RT_ARCH_ARM64)
638	uint32_t uOld;
639	# if defined(RTASM_ARM64_USE_FEAT_LSE)
640	/* SWPALH is ~40% more expensive than the non-LSE variant on an M1, 20%
641	slower if we remove the barrier. But since we have the barrier we
642	shouldn't need that, right? Ordering should be taken care of by the DMB.
643	The SWPH is rather cheap (~70% faster). */
644	__asm__ __volatile__("Lstart_ASMAtomicXchgU16_%=:\n\t"
645	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
646	"swpalh %w[uNew], %w[uOld], %[pMem]\n\t"
647	# else
648	RTASM_ARM_DMB_SY
649	"swph %w[uNew], %w[uOld], %[pMem]\n\t"
650	# endif
651	: [pMem] "+Q" (*pu16)
652	, [uOld] "=&r" (uOld)
653	: [uNew] "r" ((uint32_t)u16)
654	: );
655	# else
656	uint32_t rcSpill;
657	__asm__ __volatile__("Ltry_again_ASMAtomicXchgU16_%=:\n\t"
658	RTASM_ARM_DMB_SY
659	# if defined(RT_ARCH_ARM64)
660	"ldaxrh %w[uOld], %[pMem]\n\t"
661	"stlxrh %w[rc], %w[uNew], %[pMem]\n\t"
662	"cbnz %w[rc], Ltry_again_ASMAtomicXchgU16_%=\n\t"
663	# else
664	"ldrexh %[uOld], %[pMem]\n\t" /* ARMv6+ */
665	"strexh %[rc], %[uNew], %[pMem]\n\t"
666	"cmp %[rc], #0\n\t"
667	"bne Ltry_again_ASMAtomicXchgU16_%=\n\t"
668	# endif
669	: [pMem] "+Q" (*pu16)
670	, [uOld] "=&r" (uOld)
671	, [rc] "=&r" (rcSpill)
672	: [uNew] "r" ((uint32_t)u16)
673	RTASM_ARM_DMB_SY_COMMA_IN_REG
674	: "cc");
675	# endif
676	return (uint16_t)uOld;
677
678	# else
679	# error "Port me"
680	# endif
681	}
682	#endif
683
684
685	/**
686	* Atomically Exchange a signed 16-bit value, ordered.
687	*
688	* @returns Current *pu16 value
689	* @param pi16 Pointer to the 16-bit variable to update.
690	* @param i16 The 16-bit value to assign to *pi16.
691	*/
692	DECLINLINE(int16_t) ASMAtomicXchgS16(volatile int16_t RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF
693	{
694	return (int16_t)ASMAtomicXchgU16((volatile uint16_t RT_FAR *)pi16, (uint16_t)i16);
695	}
696
697
698	/**
699	* Atomically Exchange an unsigned 32-bit value, ordered.
700	*
701	* @returns Current *pu32 value
702	* @param pu32 Pointer to the 32-bit variable to update.
703	* @param u32 The 32-bit value to assign to *pu32.
704	*
705	* @remarks Does not work on 286 and earlier.
706	*/
707	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
708	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicXchgU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
709	#else
710	DECLINLINE(uint32_t) ASMAtomicXchgU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
711	{
712	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
713	# if RT_INLINE_ASM_GNU_STYLE
714	__asm__ __volatile__("xchgl %0, %1\n\t"
715	: "=m" (pu32) /* @todo r=bird: +m rather than =m here? */
716	, "=r" (u32)
717	: "1" (u32)
718	, "m" (*pu32));
719
720	# elif RT_INLINE_ASM_USES_INTRIN
721	u32 = _InterlockedExchange((long RT_FAR *)pu32, u32);
722
723	# else
724	__asm
725	{
726	# ifdef RT_ARCH_AMD64
727	mov rdx, [pu32]
728	mov eax, u32
729	xchg [rdx], eax
730	mov [u32], eax
731	# else
732	mov edx, [pu32]
733	mov eax, u32
734	xchg [edx], eax
735	mov [u32], eax
736	# endif
737	}
738	# endif
739	return u32;
740
741	# elif defined(RT_ARCH_ARM32) \|\| defined(RT_ARCH_ARM64)
742	uint32_t uOld;
743	# if defined(RTASM_ARM64_USE_FEAT_LSE)
744	/* SWPAL is ~40% more expensive than the non-LSE variant on an M1, 20%
745	slower if we remove the barrier. But since we have the barrier we
746	shouldn't need that, right? Ordering should be taken care of by the DMB.
747	The SWP is rather cheap (~70% faster). */
748	__asm__ __volatile__("Lstart_ASMAtomicXchgU32_%=:\n\t"
749	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
750	"swpal %w[uNew], %w[uOld], %[pMem]\n\t"
751	# else
752	RTASM_ARM_DMB_SY
753	"swp %w[uNew], %w[uOld], %[pMem]\n\t"
754	# endif
755	: [pMem] "+Q" (*pu32)
756	, [uOld] "=&r" (uOld)
757	: [uNew] "r" (u32)
758	: );
759	# else
760	uint32_t rcSpill;
761	__asm__ __volatile__("Ltry_again_ASMAtomicXchgU32_%=:\n\t"
762	RTASM_ARM_DMB_SY
763	# if defined(RT_ARCH_ARM64)
764	"ldaxr %w[uOld], %[pMem]\n\t"
765	"stlxr %w[rc], %w[uNew], %[pMem]\n\t"
766	"cbnz %w[rc], Ltry_again_ASMAtomicXchgU32_%=\n\t"
767	# else
768	"ldrex %[uOld], %[pMem]\n\t" /* ARMv6+ */
769	"strex %[rc], %[uNew], %[pMem]\n\t"
770	"cmp %[rc], #0\n\t"
771	"bne Ltry_again_ASMAtomicXchgU32_%=\n\t"
772	# endif
773	: [pMem] "+Q" (*pu32)
774	, [uOld] "=&r" (uOld)
775	, [rc] "=&r" (rcSpill)
776	: [uNew] "r" (u32)
777	RTASM_ARM_DMB_SY_COMMA_IN_REG
778	: "cc");
779	# endif
780	return uOld;
781
782	# else
783	# error "Port me"
784	# endif
785	}
786	#endif
787
788
789	/**
790	* Atomically Exchange a signed 32-bit value, ordered.
791	*
792	* @returns Current *pu32 value
793	* @param pi32 Pointer to the 32-bit variable to update.
794	* @param i32 The 32-bit value to assign to *pi32.
795	*/
796	DECLINLINE(int32_t) ASMAtomicXchgS32(volatile int32_t RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
797	{
798	return (int32_t)ASMAtomicXchgU32((volatile uint32_t RT_FAR *)pi32, (uint32_t)i32);
799	}
800
801
802	/**
803	* Atomically Exchange an unsigned 64-bit value, ordered.
804	*
805	* @returns Current *pu64 value
806	* @param pu64 Pointer to the 64-bit variable to update.
807	* @param u64 The 64-bit value to assign to *pu64.
808	*
809	* @remarks Works on 32-bit x86 CPUs starting with Pentium.
810	*/
811	#if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \
812	\|\| RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
813	RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicXchgU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
814	#else
815	DECLINLINE(uint64_t) ASMAtomicXchgU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
816	{
817	# if defined(RT_ARCH_AMD64)
818	# if RT_INLINE_ASM_USES_INTRIN
819	return _InterlockedExchange64((__int64 *)pu64, u64);
820
821	# elif RT_INLINE_ASM_GNU_STYLE
822	__asm__ __volatile__("xchgq %0, %1\n\t"
823	: "=m" (*pu64)
824	, "=r" (u64)
825	: "1" (u64)
826	, "m" (*pu64));
827	return u64;
828	# else
829	__asm
830	{
831	mov rdx, [pu64]
832	mov rax, [u64]
833	xchg [rdx], rax
834	mov [u64], rax
835	}
836	return u64;
837	# endif
838
839	# elif defined(RT_ARCH_X86)
840	# if RT_INLINE_ASM_GNU_STYLE
841	# if defined(PIC) \|\| defined(__PIC__)
842	uint32_t u32EBX = (uint32_t)u64;
843	__asm__ __volatile__(/"xchgl %%esi, %5\n\t"/
844	"xchgl %%ebx, %3\n\t"
845	"1:\n\t"
846	"lock; cmpxchg8b (%5)\n\t"
847	"jnz 1b\n\t"
848	"movl %3, %%ebx\n\t"
849	/"xchgl %%esi, %5\n\t"/
850	: "=A" (u64)
851	, "=m" (*pu64)
852	: "0" (*pu64)
853	, "m" ( u32EBX )
854	, "c" ( (uint32_t)(u64 >> 32) )
855	, "S" (pu64)
856	: "cc");
857	# else /* !PIC */
858	__asm__ __volatile__("1:\n\t"
859	"lock; cmpxchg8b %1\n\t"
860	"jnz 1b\n\t"
861	: "=A" (u64)
862	, "=m" (*pu64)
863	: "0" (*pu64)
864	, "b" ( (uint32_t)u64 )
865	, "c" ( (uint32_t)(u64 >> 32) )
866	: "cc");
867	# endif
868	# else
869	__asm
870	{
871	mov ebx, dword ptr [u64]
872	mov ecx, dword ptr [u64 + 4]
873	mov edi, pu64
874	mov eax, dword ptr [edi]
875	mov edx, dword ptr [edi + 4]
876	retry:
877	lock cmpxchg8b [edi]
878	jnz retry
879	mov dword ptr [u64], eax
880	mov dword ptr [u64 + 4], edx
881	}
882	# endif
883	return u64;
884
885	# elif defined(RT_ARCH_ARM32) \|\| defined(RT_ARCH_ARM64)
886	uint64_t uOld;
887	# if defined(RTASM_ARM64_USE_FEAT_LSE)
888	/* SWPAL is ~40% more expensive than the non-LSE variant on an M1, 20%
889	slower if we remove the barrier. But since we have the barrier we
890	shouldn't need that, right? Ordering should be taken care of by the DMB.
891	The SWP is rather cheap (~70% faster). */
892	__asm__ __volatile__("Lstart_ASMAtomicXchgU64_%=:\n\t"
893	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
894	"swpal %[uNew], %[uOld], %[pMem]\n\t"
895	# else
896	RTASM_ARM_DMB_SY
897	"swp %[uNew], %[uOld], %[pMem]\n\t"
898	# endif
899	: [pMem] "+Q" (*pu64)
900	, [uOld] "=&r" (uOld)
901	: [uNew] "r" (u64)
902	: );
903	# else
904	uint32_t rcSpill;
905	__asm__ __volatile__("Ltry_again_ASMAtomicXchgU64_%=:\n\t"
906	RTASM_ARM_DMB_SY
907	# if defined(RT_ARCH_ARM64)
908	"ldaxr %[uOld], %[pMem]\n\t"
909	"stlxr %w[rc], %[uNew], %[pMem]\n\t"
910	"cbnz %w[rc], Ltry_again_ASMAtomicXchgU64_%=\n\t"
911	# else
912	"ldrexd %[uOld], %H[uOld], %[pMem]\n\t" /* ARMv6+ */
913	"strexd %[rc], %[uNew], %H[uNew], %[pMem]\n\t"
914	"cmp %[rc], #0\n\t"
915	"bne Ltry_again_ASMAtomicXchgU64_%=\n\t"
916	# endif
917	: [pMem] "+Q" (*pu64)
918	, [uOld] "=&r" (uOld)
919	, [rc] "=&r" (rcSpill)
920	: [uNew] "r" (u64)
921	RTASM_ARM_DMB_SY_COMMA_IN_REG
922	: "cc");
923	# endif
924	return uOld;
925
926	# else
927	# error "Port me"
928	# endif
929	}
930	#endif
931
932
933	/**
934	* Atomically Exchange an signed 64-bit value, ordered.
935	*
936	* @returns Current *pi64 value
937	* @param pi64 Pointer to the 64-bit variable to update.
938	* @param i64 The 64-bit value to assign to *pi64.
939	*/
940	DECLINLINE(int64_t) ASMAtomicXchgS64(volatile int64_t RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
941	{
942	return (int64_t)ASMAtomicXchgU64((volatile uint64_t RT_FAR *)pi64, (uint64_t)i64);
943	}
944
945
946	/**
947	* Atomically Exchange a size_t value, ordered.
948	*
949	* @returns Current *ppv value
950	* @param puDst Pointer to the size_t variable to update.
951	* @param uNew The new value to assign to *puDst.
952	*/
953	DECLINLINE(size_t) ASMAtomicXchgZ(size_t volatile RT_FAR *puDst, const size_t uNew) RT_NOTHROW_DEF
954	{
955	#if ARCH_BITS == 16
956	AssertCompile(sizeof(size_t) == 2);
957	return ASMAtomicXchgU16((volatile uint16_t RT_FAR *)puDst, uNew);
958	#elif ARCH_BITS == 32
959	return ASMAtomicXchgU32((volatile uint32_t RT_FAR *)puDst, uNew);
960	#elif ARCH_BITS == 64
961	return ASMAtomicXchgU64((volatile uint64_t RT_FAR *)puDst, uNew);
962	#else
963	# error "ARCH_BITS is bogus"
964	#endif
965	}
966
967
968	/**
969	* Atomically Exchange a pointer value, ordered.
970	*
971	* @returns Current *ppv value
972	* @param ppv Pointer to the pointer variable to update.
973	* @param pv The pointer value to assign to *ppv.
974	*/
975	DECLINLINE(void RT_FAR ) ASMAtomicXchgPtr(void RT_FAR volatile RT_FAR ppv, const void RT_FAR pv) RT_NOTHROW_DEF
976	{
977	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
978	return (void RT_FAR )ASMAtomicXchgU32((volatile uint32_t RT_FAR )(void RT_FAR *)ppv, (uint32_t)pv);
979	#elif ARCH_BITS == 64
980	return (void RT_FAR )ASMAtomicXchgU64((volatile uint64_t RT_FAR )(void RT_FAR *)ppv, (uint64_t)pv);
981	#else
982	# error "ARCH_BITS is bogus"
983	#endif
984	}
985
986
987	/**
988	* Convenience macro for avoiding the annoying casting with ASMAtomicXchgPtr.
989	*
990	* @returns Current *pv value
991	* @param ppv Pointer to the pointer variable to update.
992	* @param pv The pointer value to assign to *ppv.
993	* @param Type The type of *ppv, sans volatile.
994	*/
995	#ifdef __GNUC__ /* 8.2.0 requires -Wno-ignored-qualifiers */
996	# define ASMAtomicXchgPtrT(ppv, pv, Type) \
997	__extension__ \
998	({\
999	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
1000	Type const pvTypeChecked = (pv); \
1001	Type pvTypeCheckedRet = (__typeof__((ppv))) ASMAtomicXchgPtr((void volatile )ppvTypeChecked, (void )pvTypeChecked); \
1002	pvTypeCheckedRet; \
1003	})
1004	#else
1005	# define ASMAtomicXchgPtrT(ppv, pv, Type) \
1006	(Type)ASMAtomicXchgPtr((void RT_FAR * volatile RT_FAR )(ppv), (void RT_FAR )(pv))
1007	#endif
1008
1009
1010	/**
1011	* Atomically Exchange a raw-mode context pointer value, ordered.
1012	*
1013	* @returns Current *ppv value
1014	* @param ppvRC Pointer to the pointer variable to update.
1015	* @param pvRC The pointer value to assign to *ppv.
1016	*/
1017	DECLINLINE(RTRCPTR) ASMAtomicXchgRCPtr(RTRCPTR volatile RT_FAR *ppvRC, RTRCPTR pvRC) RT_NOTHROW_DEF
1018	{
1019	return (RTRCPTR)ASMAtomicXchgU32((uint32_t volatile RT_FAR )(void RT_FAR )ppvRC, (uint32_t)pvRC);
1020	}
1021
1022
1023	/**
1024	* Atomically Exchange a ring-0 pointer value, ordered.
1025	*
1026	* @returns Current *ppv value
1027	* @param ppvR0 Pointer to the pointer variable to update.
1028	* @param pvR0 The pointer value to assign to *ppv.
1029	*/
1030	DECLINLINE(RTR0PTR) ASMAtomicXchgR0Ptr(RTR0PTR volatile RT_FAR *ppvR0, RTR0PTR pvR0) RT_NOTHROW_DEF
1031	{
1032	#if R0_ARCH_BITS == 32 \|\| ARCH_BITS == 16
1033	return (RTR0PTR)ASMAtomicXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )ppvR0, (uint32_t)pvR0);
1034	#elif R0_ARCH_BITS == 64
1035	return (RTR0PTR)ASMAtomicXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )ppvR0, (uint64_t)pvR0);
1036	#else
1037	# error "R0_ARCH_BITS is bogus"
1038	#endif
1039	}
1040
1041
1042	/**
1043	* Atomically Exchange a ring-3 pointer value, ordered.
1044	*
1045	* @returns Current *ppv value
1046	* @param ppvR3 Pointer to the pointer variable to update.
1047	* @param pvR3 The pointer value to assign to *ppv.
1048	*/
1049	DECLINLINE(RTR3PTR) ASMAtomicXchgR3Ptr(RTR3PTR volatile RT_FAR *ppvR3, RTR3PTR pvR3) RT_NOTHROW_DEF
1050	{
1051	#if R3_ARCH_BITS == 32 \|\| ARCH_BITS == 16
1052	return (RTR3PTR)ASMAtomicXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )ppvR3, (uint32_t)pvR3);
1053	#elif R3_ARCH_BITS == 64
1054	return (RTR3PTR)ASMAtomicXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )ppvR3, (uint64_t)pvR3);
1055	#else
1056	# error "R3_ARCH_BITS is bogus"
1057	#endif
1058	}
1059
1060
1061	/** @def ASMAtomicXchgHandle
1062	* Atomically Exchange a typical IPRT handle value, ordered.
1063	*
1064	* @param ph Pointer to the value to update.
1065	* @param hNew The new value to assigned to *pu.
1066	* @param phRes Where to store the current *ph value.
1067	*
1068	* @remarks This doesn't currently work for all handles (like RTFILE).
1069	*/
1070	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
1071	# define ASMAtomicXchgHandle(ph, hNew, phRes) \
1072	do { \
1073	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
1074	AssertCompile(sizeof(*(phRes)) == sizeof(uint32_t)); \
1075	(uint32_t RT_FAR )(phRes) = ASMAtomicXchgU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)(hNew)); \
1076	} while (0)
1077	#elif HC_ARCH_BITS == 64
1078	# define ASMAtomicXchgHandle(ph, hNew, phRes) \
1079	do { \
1080	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
1081	AssertCompile(sizeof(*(phRes)) == sizeof(uint64_t)); \
1082	(uint64_t RT_FAR )(phRes) = ASMAtomicXchgU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)(hNew)); \
1083	} while (0)
1084	#else
1085	# error HC_ARCH_BITS
1086	#endif
1087
1088
1089	/**
1090	* Atomically Exchange a value which size might differ
1091	* between platforms or compilers, ordered.
1092	*
1093	* @param pu Pointer to the variable to update.
1094	* @param uNew The value to assign to *pu.
1095	* @todo This is busted as its missing the result argument.
1096	*/
1097	#define ASMAtomicXchgSize(pu, uNew) \
1098	do { \
1099	switch (sizeof(*(pu))) { \
1100	case 1: ASMAtomicXchgU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu), (uint8_t)(uNew)); break; \
1101	case 2: ASMAtomicXchgU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu), (uint16_t)(uNew)); break; \
1102	case 4: ASMAtomicXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
1103	case 8: ASMAtomicXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
1104	default: AssertMsgFailed(("ASMAtomicXchgSize: size %d is not supported\n", sizeof(*(pu)))); \
1105	} \
1106	} while (0)
1107
1108	/**
1109	* Atomically Exchange a value which size might differ
1110	* between platforms or compilers, ordered.
1111	*
1112	* @param pu Pointer to the variable to update.
1113	* @param uNew The value to assign to *pu.
1114	* @param puRes Where to store the current *pu value.
1115	*/
1116	#define ASMAtomicXchgSizeCorrect(pu, uNew, puRes) \
1117	do { \
1118	switch (sizeof(*(pu))) { \
1119	case 1: (uint8_t RT_FAR )(puRes) = ASMAtomicXchgU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu), (uint8_t)(uNew)); break; \
1120	case 2: (uint16_t RT_FAR )(puRes) = ASMAtomicXchgU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu), (uint16_t)(uNew)); break; \
1121	case 4: (uint32_t RT_FAR )(puRes) = ASMAtomicXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
1122	case 8: (uint64_t RT_FAR )(puRes) = ASMAtomicXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
1123	default: AssertMsgFailed(("ASMAtomicXchgSize: size %d is not supported\n", sizeof(*(pu)))); \
1124	} \
1125	} while (0)
1126
1127
1128
1129	/**
1130	* Atomically Compare and Exchange an unsigned 8-bit value, ordered.
1131	*
1132	* @returns true if xchg was done.
1133	* @returns false if xchg wasn't done.
1134	*
1135	* @param pu8 Pointer to the value to update.
1136	* @param u8New The new value to assigned to *pu8.
1137	* @param u8Old The old value to *pu8 compare with.
1138	*
1139	* @remarks x86: Requires a 486 or later.
1140	* @todo Rename ASMAtomicCmpWriteU8
1141	*/
1142	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM \|\| !RT_INLINE_ASM_GNU_STYLE
1143	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgU8(volatile uint8_t RT_FAR *pu8, const uint8_t u8New, const uint8_t u8Old) RT_NOTHROW_PROTO;
1144	#else
1145	DECLINLINE(bool) ASMAtomicCmpXchgU8(volatile uint8_t RT_FAR *pu8, const uint8_t u8New, uint8_t u8Old) RT_NOTHROW_DEF
1146	{
1147	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1148	uint8_t u8Ret;
1149	__asm__ __volatile__("lock; cmpxchgb %3, %0\n\t"
1150	"setz %1\n\t"
1151	: "=m" (*pu8)
1152	, "=qm" (u8Ret)
1153	, "=a" (u8Old)
1154	: "q" (u8New)
1155	, "2" (u8Old)
1156	, "m" (*pu8)
1157	: "cc");
1158	return (bool)u8Ret;
1159
1160	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
1161	union { uint32_t u; bool f; } fXchg;
1162	uint32_t u32Spill;
1163	# if defined(RTASM_ARM64_USE_FEAT_LSE)
1164	__asm__ __volatile__("Lstart_ASMAtomicCmpXchgU8_%=:\n\t"
1165	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) /* M1 bench: casalb=5625 vs dmb+casb=1597 vs non-lse=5623 (ps/call) */
1166	"casalb %w[uOldActual], %w[uNew], %[pMem]\n\t"
1167	# else
1168	RTASM_ARM_DMB_SY
1169	"casb %w[uOldActual], %w[uNew], %[pMem]\n\t"
1170	# endif
1171	"cmp %w[uOldActual], %w[uOldOrg]\n\t"
1172	"cset %w[fXchg], eq\n\t"
1173	: [pMem] "+Q" (*pu8)
1174	, [uOldActual] "=&r" (u32Spill)
1175	, [fXchg] "=&r" (fXchg.u)
1176	: [uNew] "r" ((uint32_t)u8New)
1177	, [uOldOrg] "r" ((uint32_t)u8Old)
1178	, "[uOldActual]" ((uint32_t)u8Old)
1179	: "cc");
1180	# else
1181	uint32_t rcSpill;
1182	__asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgU8_%=:\n\t"
1183	RTASM_ARM_DMB_SY
1184	# if defined(RT_ARCH_ARM64)
1185	"ldaxrb %w[uOld], %[pMem]\n\t"
1186	"cmp %w[uOld], %w[uCmp]\n\t"
1187	"bne 1f\n\t" /* stop here if not equal */
1188	"stlxrb %w[rc], %w[uNew], %[pMem]\n\t"
1189	"cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgU8_%=\n\t"
1190	"mov %w[fXchg], #1\n\t"
1191	"1:\n\t"
1192	"clrex\n\t"
1193	# else
1194	"ldrexb %[uOld], %[pMem]\n\t"
1195	"teq %[uOld], %[uCmp]\n\t"
1196	"strexbeq %[rc], %[uNew], %[pMem]\n\t"
1197	"bne 1f\n\t" /* stop here if not equal */
1198	"cmp %[rc], #0\n\t"
1199	"bne Ltry_again_ASMAtomicCmpXchgU8_%=\n\t"
1200	"mov %[fXchg], #1\n\t"
1201	"1:\n\t"
1202	/** @todo clrexne on armv7? */
1203	# endif
1204	: [pMem] "+Q" (*pu8)
1205	, [uOld] "=&r" (u32Spill)
1206	, [rc] "=&r" (rcSpill)
1207	, [fXchg] "=&r" (fXchg.u)
1208	: [uCmp] "r" ((uint32_t)u8Old)
1209	, [uNew] "r" ((uint32_t)u8New)
1210	, "[fXchg]" (0)
1211	RTASM_ARM_DMB_SY_COMMA_IN_REG
1212	: "cc");
1213	# endif
1214	return fXchg.f;
1215
1216	# else
1217	# error "Port me"
1218	# endif
1219	}
1220	#endif
1221
1222
1223	/**
1224	* Atomically Compare and Exchange a signed 8-bit value, ordered.
1225	*
1226	* @returns true if xchg was done.
1227	* @returns false if xchg wasn't done.
1228	*
1229	* @param pi8 Pointer to the value to update.
1230	* @param i8New The new value to assigned to *pi8.
1231	* @param i8Old The old value to *pi8 compare with.
1232	*
1233	* @remarks x86: Requires a 486 or later.
1234	* @todo Rename ASMAtomicCmpWriteS8
1235	*/
1236	DECLINLINE(bool) ASMAtomicCmpXchgS8(volatile int8_t RT_FAR *pi8, const int8_t i8New, const int8_t i8Old) RT_NOTHROW_DEF
1237	{
1238	return ASMAtomicCmpXchgU8((volatile uint8_t RT_FAR *)pi8, (uint8_t)i8New, (uint8_t)i8Old);
1239	}
1240
1241
1242	/**
1243	* Atomically Compare and Exchange a bool value, ordered.
1244	*
1245	* @returns true if xchg was done.
1246	* @returns false if xchg wasn't done.
1247	*
1248	* @param pf Pointer to the value to update.
1249	* @param fNew The new value to assigned to *pf.
1250	* @param fOld The old value to *pf compare with.
1251	*
1252	* @remarks x86: Requires a 486 or later.
1253	* @todo Rename ASMAtomicCmpWriteBool
1254	*/
1255	DECLINLINE(bool) ASMAtomicCmpXchgBool(volatile bool RT_FAR *pf, const bool fNew, const bool fOld) RT_NOTHROW_DEF
1256	{
1257	return ASMAtomicCmpXchgU8((volatile uint8_t RT_FAR *)pf, (uint8_t)fNew, (uint8_t)fOld);
1258	}
1259
1260
1261	/**
1262	* Atomically Compare and Exchange an unsigned 32-bit value, ordered.
1263	*
1264	* @returns true if xchg was done.
1265	* @returns false if xchg wasn't done.
1266	*
1267	* @param pu32 Pointer to the value to update.
1268	* @param u32New The new value to assigned to *pu32.
1269	* @param u32Old The old value to *pu32 compare with.
1270	*
1271	* @remarks x86: Requires a 486 or later.
1272	* @todo Rename ASMAtomicCmpWriteU32
1273	*/
1274	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
1275	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgU32(volatile uint32_t RT_FAR *pu32, const uint32_t u32New, const uint32_t u32Old) RT_NOTHROW_PROTO;
1276	#else
1277	DECLINLINE(bool) ASMAtomicCmpXchgU32(volatile uint32_t RT_FAR *pu32, const uint32_t u32New, uint32_t u32Old) RT_NOTHROW_DEF
1278	{
1279	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1280	# if RT_INLINE_ASM_GNU_STYLE
1281	uint8_t u8Ret;
1282	__asm__ __volatile__("lock; cmpxchgl %3, %0\n\t"
1283	"setz %1\n\t"
1284	: "=m" (*pu32)
1285	, "=qm" (u8Ret)
1286	, "=a" (u32Old)
1287	: "r" (u32New)
1288	, "2" (u32Old)
1289	, "m" (*pu32)
1290	: "cc");
1291	return (bool)u8Ret;
1292
1293	# elif RT_INLINE_ASM_USES_INTRIN
1294	return (uint32_t)_InterlockedCompareExchange((long RT_FAR *)pu32, u32New, u32Old) == u32Old;
1295
1296	# else
1297	uint32_t u32Ret;
1298	__asm
1299	{
1300	# ifdef RT_ARCH_AMD64
1301	mov rdx, [pu32]
1302	# else
1303	mov edx, [pu32]
1304	# endif
1305	mov eax, [u32Old]
1306	mov ecx, [u32New]
1307	# ifdef RT_ARCH_AMD64
1308	lock cmpxchg [rdx], ecx
1309	# else
1310	lock cmpxchg [edx], ecx
1311	# endif
1312	setz al
1313	movzx eax, al
1314	mov [u32Ret], eax
1315	}
1316	return !!u32Ret;
1317	# endif
1318
1319	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
1320	union { uint32_t u; bool f; } fXchg;
1321	uint32_t u32Spill;
1322	/* M1 bench: match: casal= 6592 vs dmb+cas= 1562 vs non-lse=5634 (ps/call)
1323	mismatch: casal=18794 vs dmb+cas=19697 vs non-lse=2499 (ps/call) */
1324	# if defined(RTASM_ARM64_USE_FEAT_LSE)
1325	__asm__ __volatile__("Lstart_ASMAtomicCmpXchgU32_%=:\n\t"
1326	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
1327	"casal %w[uOldActual], %w[uNew], %[pMem]\n\t"
1328	# else
1329	RTASM_ARM_DMB_SY
1330	"cas %w[uOldActual], %w[uNew], %[pMem]\n\t"
1331	# endif
1332	"cmp %w[uOldActual], %w[uOldOrg]\n\t"
1333	"cset %w[fXchg], eq\n\t"
1334	: [pMem] "+Q" (*pu32)
1335	, [uOldActual] "=&r" (u32Spill)
1336	, [fXchg] "=&r" (fXchg.u)
1337	: [uNew] "r" (u32New)
1338	, [uOldOrg] "r" (u32Old)
1339	, "[uOldActual]" (u32Old)
1340	: "cc");
1341	# else
1342	uint32_t rcSpill;
1343	__asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgU32_%=:\n\t"
1344	RTASM_ARM_DMB_SY
1345	# if defined(RT_ARCH_ARM64)
1346	"ldaxr %w[uOld], %[pMem]\n\t"
1347	"cmp %w[uOld], %w[uCmp]\n\t"
1348	"bne 1f\n\t" /* stop here if not equal */
1349	"stlxr %w[rc], %w[uNew], %[pMem]\n\t"
1350	"cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgU32_%=\n\t"
1351	"mov %w[fXchg], #1\n\t"
1352	"1:\n\t"
1353	"clrex\n\t"
1354	# else
1355	"ldrex %[uOld], %[pMem]\n\t"
1356	"teq %[uOld], %[uCmp]\n\t"
1357	"strexeq %[rc], %[uNew], %[pMem]\n\t"
1358	"bne 1f\n\t" /* stop here if not equal */
1359	"cmp %[rc], #0\n\t"
1360	"bne Ltry_again_ASMAtomicCmpXchgU32_%=\n\t"
1361	"mov %[fXchg], #1\n\t"
1362	"1:\n\t"
1363	/** @todo clrexne on armv7? */
1364	# endif
1365	: [pMem] "+Q" (*pu32)
1366	, [uOld] "=&r" (u32Spill)
1367	, [rc] "=&r" (rcSpill)
1368	, [fXchg] "=&r" (fXchg.u)
1369	: [uCmp] "r" (u32Old)
1370	, [uNew] "r" (u32New)
1371	, "[fXchg]" (0)
1372	RTASM_ARM_DMB_SY_COMMA_IN_REG
1373	: "cc");
1374	# endif
1375	return fXchg.f;
1376
1377	# else
1378	# error "Port me"
1379	# endif
1380	}
1381	#endif
1382
1383
1384	/**
1385	* Atomically Compare and Exchange a signed 32-bit value, ordered.
1386	*
1387	* @returns true if xchg was done.
1388	* @returns false if xchg wasn't done.
1389	*
1390	* @param pi32 Pointer to the value to update.
1391	* @param i32New The new value to assigned to *pi32.
1392	* @param i32Old The old value to *pi32 compare with.
1393	*
1394	* @remarks x86: Requires a 486 or later.
1395	* @todo Rename ASMAtomicCmpWriteS32
1396	*/
1397	DECLINLINE(bool) ASMAtomicCmpXchgS32(volatile int32_t RT_FAR *pi32, const int32_t i32New, const int32_t i32Old) RT_NOTHROW_DEF
1398	{
1399	return ASMAtomicCmpXchgU32((volatile uint32_t RT_FAR *)pi32, (uint32_t)i32New, (uint32_t)i32Old);
1400	}
1401
1402
1403	/**
1404	* Atomically Compare and exchange an unsigned 64-bit value, ordered.
1405	*
1406	* @returns true if xchg was done.
1407	* @returns false if xchg wasn't done.
1408	*
1409	* @param pu64 Pointer to the 64-bit variable to update.
1410	* @param u64New The 64-bit value to assign to *pu64.
1411	* @param u64Old The value to compare with.
1412	*
1413	* @remarks x86: Requires a Pentium or later.
1414	* @todo Rename ASMAtomicCmpWriteU64
1415	*/
1416	#if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \
1417	\|\| RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
1418	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgU64(volatile uint64_t RT_FAR *pu64, const uint64_t u64New, const uint64_t u64Old) RT_NOTHROW_PROTO;
1419	#else
1420	DECLINLINE(bool) ASMAtomicCmpXchgU64(volatile uint64_t RT_FAR *pu64, uint64_t u64New, uint64_t u64Old) RT_NOTHROW_DEF
1421	{
1422	# if RT_INLINE_ASM_USES_INTRIN
1423	return (uint64_t)_InterlockedCompareExchange64((__int64 RT_FAR *)pu64, u64New, u64Old) == u64Old;
1424
1425	# elif defined(RT_ARCH_AMD64)
1426	# if RT_INLINE_ASM_GNU_STYLE
1427	uint8_t u8Ret;
1428	__asm__ __volatile__("lock; cmpxchgq %3, %0\n\t"
1429	"setz %1\n\t"
1430	: "=m" (*pu64)
1431	, "=qm" (u8Ret)
1432	, "=a" (u64Old)
1433	: "r" (u64New)
1434	, "2" (u64Old)
1435	, "m" (*pu64)
1436	: "cc");
1437	return (bool)u8Ret;
1438	# else
1439	bool fRet;
1440	__asm
1441	{
1442	mov rdx, [pu32]
1443	mov rax, [u64Old]
1444	mov rcx, [u64New]
1445	lock cmpxchg [rdx], rcx
1446	setz al
1447	mov [fRet], al
1448	}
1449	return fRet;
1450	# endif
1451
1452	# elif defined(RT_ARCH_X86)
1453	uint32_t u32Ret;
1454	# if RT_INLINE_ASM_GNU_STYLE
1455	# if defined(PIC) \|\| defined(__PIC__)
1456	uint32_t u32EBX = (uint32_t)u64New;
1457	uint32_t u32Spill;
1458	__asm__ __volatile__("xchgl %%ebx, %4\n\t"
1459	"lock; cmpxchg8b (%6)\n\t"
1460	"setz %%al\n\t"
1461	"movl %4, %%ebx\n\t"
1462	"movzbl %%al, %%eax\n\t"
1463	: "=a" (u32Ret)
1464	, "=d" (u32Spill)
1465	# if RT_GNUC_PREREQ(4, 3)
1466	, "+m" (*pu64)
1467	# else
1468	, "=m" (*pu64)
1469	# endif
1470	: "A" (u64Old)
1471	, "m" ( u32EBX )
1472	, "c" ( (uint32_t)(u64New >> 32) )
1473	, "S" (pu64)
1474	: "cc");
1475	# else /* !PIC */
1476	uint32_t u32Spill;
1477	__asm__ __volatile__("lock; cmpxchg8b %2\n\t"
1478	"setz %%al\n\t"
1479	"movzbl %%al, %%eax\n\t"
1480	: "=a" (u32Ret)
1481	, "=d" (u32Spill)
1482	, "+m" (*pu64)
1483	: "A" (u64Old)
1484	, "b" ( (uint32_t)u64New )
1485	, "c" ( (uint32_t)(u64New >> 32) )
1486	: "cc");
1487	# endif
1488	return (bool)u32Ret;
1489	# else
1490	__asm
1491	{
1492	mov ebx, dword ptr [u64New]
1493	mov ecx, dword ptr [u64New + 4]
1494	mov edi, [pu64]
1495	mov eax, dword ptr [u64Old]
1496	mov edx, dword ptr [u64Old + 4]
1497	lock cmpxchg8b [edi]
1498	setz al
1499	movzx eax, al
1500	mov dword ptr [u32Ret], eax
1501	}
1502	return !!u32Ret;
1503	# endif
1504
1505	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
1506	union { uint32_t u; bool f; } fXchg;
1507	uint64_t u64Spill;
1508	/* M1 bench: match: casal= 6599 vs dmb+cas= 1565 vs non-lse=5000 (ps/call)
1509	mismatch: casal=18797 vs dmb+cas=19731 vs non-lse=2512 (ps/call) */
1510	# if defined(RTASM_ARM64_USE_FEAT_LSE)
1511	__asm__ __volatile__("Lstart_ASMAtomicCmpXchgU75_%=:\n\t"
1512	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
1513	"casal %[uOldActual], %[uNew], %[pMem]\n\t"
1514	# else
1515	RTASM_ARM_DMB_SY
1516	"cas %[uOldActual], %[uNew], %[pMem]\n\t"
1517	# endif
1518	"cmp %[uOldActual], %[uOldOrg]\n\t"
1519	"cset %w[fXchg], eq\n\t"
1520	: [pMem] "+Q" (*pu64)
1521	, [uOldActual] "=&r" (u64Spill)
1522	, [fXchg] "=&r" (fXchg.u)
1523	: [uNew] "r" (u64New)
1524	, [uOldOrg] "r" (u64Old)
1525	, "[uOldActual]" (u64Old)
1526	: "cc");
1527	# else
1528	uint32_t rcSpill;
1529	__asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgU64_%=:\n\t"
1530	RTASM_ARM_DMB_SY
1531	# if defined(RT_ARCH_ARM64)
1532	"ldaxr %[uOld], %[pMem]\n\t"
1533	"cmp %[uOld], %[uCmp]\n\t"
1534	"bne 1f\n\t" /* stop here if not equal */
1535	"stlxr %w[rc], %[uNew], %[pMem]\n\t"
1536	"cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
1537	"mov %w[fXchg], #1\n\t"
1538	"1:\n\t"
1539	"clrex\n\t"
1540	# else
1541	"ldrexd %[uOld], %H[uOld], %[pMem]\n\t"
1542	"teq %[uOld], %[uCmp]\n\t"
1543	"teqeq %H[uOld], %H[uCmp]\n\t"
1544	"strexdeq %[rc], %[uNew], %H[uNew], %[pMem]\n\t"
1545	"bne 1f\n\t" /* stop here if not equal */
1546	"cmp %[rc], #0\n\t"
1547	"bne Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
1548	"mov %[fXchg], #1\n\t"
1549	"1:\n\t"
1550	/** @todo clrexne on armv7? */
1551	# endif
1552	: [pMem] "+Q" (*pu64)
1553	, [uOld] "=&r" (u64Spill)
1554	, [rc] "=&r" (rcSpill)
1555	, [fXchg] "=&r" (fXchg.u)
1556	: [uCmp] "r" (u64Old)
1557	, [uNew] "r" (u64New)
1558	, "[fXchg]" (0)
1559	RTASM_ARM_DMB_SY_COMMA_IN_REG
1560	: "cc");
1561	# endif
1562	return fXchg.f;
1563
1564	# else
1565	# error "Port me"
1566	# endif
1567	}
1568	#endif
1569
1570
1571	/**
1572	* Atomically Compare and exchange a signed 64-bit value, ordered.
1573	*
1574	* @returns true if xchg was done.
1575	* @returns false if xchg wasn't done.
1576	*
1577	* @param pi64 Pointer to the 64-bit variable to update.
1578	* @param i64 The 64-bit value to assign to *pu64.
1579	* @param i64Old The value to compare with.
1580	*
1581	* @remarks x86: Requires a Pentium or later.
1582	* @todo Rename ASMAtomicCmpWriteS64
1583	*/
1584	DECLINLINE(bool) ASMAtomicCmpXchgS64(volatile int64_t RT_FAR *pi64, const int64_t i64, const int64_t i64Old) RT_NOTHROW_DEF
1585	{
1586	return ASMAtomicCmpXchgU64((volatile uint64_t RT_FAR *)pi64, (uint64_t)i64, (uint64_t)i64Old);
1587	}
1588
1589	#if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_ARM64) \|\| defined(DOXYGEN_RUNNING)
1590
1591	/** @def RTASM_HAVE_CMP_WRITE_U128
1592	* Indicates that we've got ASMAtomicCmpWriteU128(), ASMAtomicCmpWriteU128v2()
1593	* and ASMAtomicCmpWriteExU128() available. */
1594	# define RTASM_HAVE_CMP_WRITE_U128 1
1595
1596
1597	/**
1598	* Atomically compare and write an unsigned 128-bit value, ordered.
1599	*
1600	* @returns true if write was done.
1601	* @returns false if write wasn't done.
1602	*
1603	* @param pu128 Pointer to the 128-bit variable to update.
1604	* @param u64NewHi The high 64 bits of the value to assign to *pu128.
1605	* @param u64NewLo The low 64 bits of the value to assign to *pu128.
1606	* @param u64OldHi The high 64-bit of the value to compare with.
1607	* @param u64OldLo The low 64-bit of the value to compare with.
1608	*
1609	* @remarks AMD64: Not present in the earliest CPUs, so check CPUID.
1610	*/
1611	# if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN)
1612	DECLASM(bool) ASMAtomicCmpWriteU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo,
1613	const uint64_t u64OldHi, const uint64_t u64OldLo) RT_NOTHROW_PROTO;
1614	# else
1615	DECLINLINE(bool) ASMAtomicCmpWriteU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo,
1616	const uint64_t u64OldHi, const uint64_t u64OldLo) RT_NOTHROW_DEF
1617	{
1618	# if RT_INLINE_ASM_USES_INTRIN
1619	__int64 ai64Cmp[2];
1620	ai64Cmp[0] = u64OldLo;
1621	ai64Cmp[1] = u64OldHi;
1622	return _InterlockedCompareExchange128((__int64 volatile *)pu128, u64NewHi, u64NewLo, ai64Cmp) != 0;
1623
1624	# elif (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
1625	return __sync_bool_compare_and_swap(pu128, ((uint128_t)u64OldHi << 64) \| u64OldLo, ((uint128_t)u64NewHi << 64) \| u64NewLo);
1626
1627	# elif defined(RT_ARCH_AMD64)
1628	# if RT_INLINE_ASM_GNU_STYLE
1629	uint64_t u64Ret;
1630	uint64_t u64Spill;
1631	__asm__ __volatile__("lock; cmpxchg16b %2\n\t"
1632	"setz %%al\n\t"
1633	"movzbl %%al, %%eax\n\t"
1634	: "=a" (u64Ret)
1635	, "=d" (u64Spill)
1636	, "+m" (*pu128)
1637	: "a" (u64OldLo)
1638	, "d" (u64OldHi)
1639	, "b" (u64NewLo)
1640	, "c" (u64NewHi)
1641	: "cc");
1642
1643	return (bool)u64Ret;
1644	# else
1645	# error "Port me"
1646	# endif
1647	# else
1648	# error "Port me"
1649	# endif
1650	}
1651	# endif
1652
1653
1654	/**
1655	* Atomically compare and write an unsigned 128-bit value, ordered.
1656	*
1657	* @returns true if write was done.
1658	* @returns false if write wasn't done.
1659	*
1660	* @param pu128 Pointer to the 128-bit variable to update.
1661	* @param u128New The 128-bit value to assign to *pu128.
1662	* @param u128Old The value to compare with.
1663	*
1664	* @remarks AMD64: Not present in the earliest CPUs, so check CPUID.
1665	*/
1666	DECLINLINE(bool) ASMAtomicCmpWriteU128(volatile uint128_t *pu128, const uint128_t u128New, const uint128_t u128Old) RT_NOTHROW_DEF
1667	{
1668	# ifdef RT_COMPILER_WITH_128BIT_INT_TYPES
1669	# if (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
1670	return __sync_bool_compare_and_swap(pu128, u128Old, u128New);
1671	# else
1672	return ASMAtomicCmpWriteU128v2(pu128, (uint64_t)(u128New >> 64), (uint64_t)u128New,
1673	(uint64_t)(u128Old >> 64), (uint64_t)u128Old);
1674	# endif
1675	# else
1676	return ASMAtomicCmpWriteU128v2(pu128, u128New.Hi, u128New.Lo, u128Old.Hi, u128Old.Lo);
1677	# endif
1678	}
1679
1680
1681	/**
1682	* RTUINT128U wrapper for ASMAtomicCmpWriteU128.
1683	*/
1684	DECLINLINE(bool) ASMAtomicCmpWriteU128U(volatile RTUINT128U *pu128, const RTUINT128U u128New,
1685	const RTUINT128U u128Old) RT_NOTHROW_DEF
1686	{
1687	# if (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
1688	return ASMAtomicCmpWriteU128(&pu128->u, u128New.u, u128Old.u);
1689	# else
1690	return ASMAtomicCmpWriteU128v2(&pu128->u, u128New.s.Hi, u128New.s.Lo, u128Old.s.Hi, u128Old.s.Lo);
1691	# endif
1692	}
1693
1694	#endif /* RT_ARCH_AMD64 \|\| RT_ARCH_ARM64 */
1695
1696	/**
1697	* Atomically Compare and Exchange a pointer value, ordered.
1698	*
1699	* @returns true if xchg was done.
1700	* @returns false if xchg wasn't done.
1701	*
1702	* @param ppv Pointer to the value to update.
1703	* @param pvNew The new value to assigned to *ppv.
1704	* @param pvOld The old value to *ppv compare with.
1705	*
1706	* @remarks x86: Requires a 486 or later.
1707	* @todo Rename ASMAtomicCmpWritePtrVoid
1708	*/
1709	DECLINLINE(bool) ASMAtomicCmpXchgPtrVoid(void RT_FAR * volatile RT_FAR ppv, const void RT_FAR pvNew, const void RT_FAR *pvOld) RT_NOTHROW_DEF
1710	{
1711	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
1712	return ASMAtomicCmpXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )ppv, (uint32_t)pvNew, (uint32_t)pvOld);
1713	#elif ARCH_BITS == 64
1714	return ASMAtomicCmpXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )ppv, (uint64_t)pvNew, (uint64_t)pvOld);
1715	#else
1716	# error "ARCH_BITS is bogus"
1717	#endif
1718	}
1719
1720
1721	/**
1722	* Atomically Compare and Exchange a pointer value, ordered.
1723	*
1724	* @returns true if xchg was done.
1725	* @returns false if xchg wasn't done.
1726	*
1727	* @param ppv Pointer to the value to update.
1728	* @param pvNew The new value to assigned to *ppv.
1729	* @param pvOld The old value to *ppv compare with.
1730	*
1731	* @remarks This is relatively type safe on GCC platforms.
1732	* @remarks x86: Requires a 486 or later.
1733	* @todo Rename ASMAtomicCmpWritePtr
1734	*/
1735	#ifdef __GNUC__
1736	# define ASMAtomicCmpXchgPtr(ppv, pvNew, pvOld) \
1737	__extension__ \
1738	({\
1739	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
1740	__typeof__(*(ppv)) const pvNewTypeChecked = (pvNew); \
1741	__typeof__(*(ppv)) const pvOldTypeChecked = (pvOld); \
1742	bool fMacroRet = ASMAtomicCmpXchgPtrVoid((void * volatile *)ppvTypeChecked, \
1743	(void )pvNewTypeChecked, (void )pvOldTypeChecked); \
1744	fMacroRet; \
1745	})
1746	#else
1747	# define ASMAtomicCmpXchgPtr(ppv, pvNew, pvOld) \
1748	ASMAtomicCmpXchgPtrVoid((void RT_FAR * volatile RT_FAR )(ppv), (void RT_FAR )(pvNew), (void RT_FAR *)(pvOld))
1749	#endif
1750
1751
1752	/** @def ASMAtomicCmpXchgHandle
1753	* Atomically Compare and Exchange a typical IPRT handle value, ordered.
1754	*
1755	* @param ph Pointer to the value to update.
1756	* @param hNew The new value to assigned to *pu.
1757	* @param hOld The old value to *pu compare with.
1758	* @param fRc Where to store the result.
1759	*
1760	* @remarks This doesn't currently work for all handles (like RTFILE).
1761	* @remarks x86: Requires a 486 or later.
1762	* @todo Rename ASMAtomicCmpWriteHandle
1763	*/
1764	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
1765	# define ASMAtomicCmpXchgHandle(ph, hNew, hOld, fRc) \
1766	do { \
1767	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
1768	(fRc) = ASMAtomicCmpXchgU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)(hNew), (const uint32_t)(hOld)); \
1769	} while (0)
1770	#elif HC_ARCH_BITS == 64
1771	# define ASMAtomicCmpXchgHandle(ph, hNew, hOld, fRc) \
1772	do { \
1773	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
1774	(fRc) = ASMAtomicCmpXchgU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)(hNew), (const uint64_t)(hOld)); \
1775	} while (0)
1776	#else
1777	# error HC_ARCH_BITS
1778	#endif
1779
1780
1781	/** @def ASMAtomicCmpXchgSize
1782	* Atomically Compare and Exchange a value which size might differ
1783	* between platforms or compilers, ordered.
1784	*
1785	* @param pu Pointer to the value to update.
1786	* @param uNew The new value to assigned to *pu.
1787	* @param uOld The old value to *pu compare with.
1788	* @param fRc Where to store the result.
1789	*
1790	* @remarks x86: Requires a 486 or later.
1791	* @todo Rename ASMAtomicCmpWriteSize
1792	*/
1793	#define ASMAtomicCmpXchgSize(pu, uNew, uOld, fRc) \
1794	do { \
1795	switch (sizeof(*(pu))) { \
1796	case 4: (fRc) = ASMAtomicCmpXchgU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew), (uint32_t)(uOld)); \
1797	break; \
1798	case 8: (fRc) = ASMAtomicCmpXchgU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew), (uint64_t)(uOld)); \
1799	break; \
1800	default: AssertMsgFailed(("ASMAtomicCmpXchgSize: size %d is not supported\n", sizeof(*(pu)))); \
1801	(fRc) = false; \
1802	break; \
1803	} \
1804	} while (0)
1805
1806
1807	/**
1808	* Atomically Compare and Exchange an unsigned 8-bit value, additionally passes
1809	* back old value, ordered.
1810	*
1811	* @returns true if xchg was done.
1812	* @returns false if xchg wasn't done.
1813	*
1814	* @param pu8 Pointer to the value to update.
1815	* @param u8New The new value to assigned to *pu32.
1816	* @param u8Old The old value to *pu8 compare with.
1817	* @param pu8Old Pointer store the old value at.
1818	*
1819	* @remarks x86: Requires a 486 or later.
1820	*/
1821	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
1822	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;
1823	#else
1824	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
1825	{
1826	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1827	# if RT_INLINE_ASM_GNU_STYLE
1828	uint8_t u8Ret;
1829	__asm__ __volatile__("lock; cmpxchgb %3, %0\n\t"
1830	"setz %1\n\t"
1831	: "=m" (*pu8)
1832	, "=qm" (u8Ret)
1833	, "=a" (*pu8Old)
1834	# if defined(RT_ARCH_X86)
1835	: "q" (u8New)
1836	# else
1837	: "r" (u8New)
1838	# endif
1839	, "a" (u8Old)
1840	, "m" (*pu8)
1841	: "cc");
1842	return (bool)u8Ret;
1843
1844	# elif RT_INLINE_ASM_USES_INTRIN
1845	return (pu8Old = _InterlockedCompareExchange8((char RT_FAR )pu8, u8New, u8Old)) == u8Old;
1846
1847	# else
1848	uint8_t u8Ret;
1849	__asm
1850	{
1851	# ifdef RT_ARCH_AMD64
1852	mov rdx, [pu8]
1853	# else
1854	mov edx, [pu8]
1855	# endif
1856	mov eax, [u8Old]
1857	mov ecx, [u8New]
1858	# ifdef RT_ARCH_AMD64
1859	lock cmpxchg [rdx], ecx
1860	mov rdx, [pu8Old]
1861	mov [rdx], eax
1862	# else
1863	lock cmpxchg [edx], ecx
1864	mov edx, [pu8Old]
1865	mov [edx], eax
1866	# endif
1867	setz al
1868	movzx eax, al
1869	mov [u8Ret], eax
1870	}
1871	return !!u8Ret;
1872	# endif
1873
1874	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
1875	/* M1 bench: match: casalb= 6594 vs dmb+casb= 1561 vs non-lse=5051 (ps/call)
1876	mismatch: casalb=15346 vs dmb+casb=16349 vs non-lse=2505 (ps/call) */
1877	# if defined(RTASM_ARM64_USE_FEAT_LSE)
1878	union { uint32_t u; bool f; } fXchg;
1879	uint32_t u32Actual;
1880	__asm__ __volatile__("Lstart_ASMAtomicCmpXchgExU8_%=:\n\t"
1881	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
1882	"casalb %w[uOldActual], %w[uNew], %[pMem]\n\t"
1883	# else
1884	RTASM_ARM_DMB_SY
1885	"casb %w[uOldActual], %w[uNew], %[pMem]\n\t"
1886	# endif
1887	"cmp %w[uOldActual], %w[uOldOrg]\n\t"
1888	"cset %w[fXchg], eq\n\t"
1889	: [pMem] "+Q" (*pu8)
1890	, [uOldActual] "=&r" (u32Actual)
1891	, [fXchg] "=&r" (fXchg.u)
1892	: [uNew] "r" ((uint32_t)u8New)
1893	, [uOldOrg] "r" ((uint32_t)u8Old)
1894	, "[uOldActual]" ((uint32_t)u8Old)
1895	: "cc");
1896	*pu8Old = (uint8_t)u32Actual;
1897	# else
1898	union { uint8_t u; bool f; } fXchg;
1899	uint8_t u8ActualOld;
1900	uint8_t rcSpill;
1901	__asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgExU8_%=:\n\t"
1902	RTASM_ARM_DMB_SY
1903	# if defined(RT_ARCH_ARM64)
1904	"ldaxrb %w[uOld], %[pMem]\n\t"
1905	"cmp %w[uOld], %w[uCmp]\n\t"
1906	"bne 1f\n\t" /* stop here if not equal */
1907	"stlxrb %w[rc], %w[uNew], %[pMem]\n\t"
1908	"cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgExU8_%=\n\t"
1909	"mov %w[fXchg], #1\n\t"
1910	"1:\n\t"
1911	"clrex\n\t"
1912	# else
1913	"ldrexb %[uOld], %[pMem]\n\t"
1914	"teq %[uOld], %[uCmp]\n\t"
1915	"strexbeq %[rc], %[uNew], %[pMem]\n\t"
1916	"bne 1f\n\t" /* stop here if not equal */
1917	"cmp %[rc], #0\n\t"
1918	"bne Ltry_again_ASMAtomicCmpXchgExU8_%=\n\t"
1919	"mov %[fXchg], #1\n\t"
1920	"1:\n\t"
1921	/** @todo clrexne on armv7? */
1922	# endif
1923	: [pMem] "+Q" (*pu8)
1924	, [uOld] "=&r" (u8ActualOld)
1925	, [rc] "=&r" (rcSpill)
1926	, [fXchg] "=&r" (fXchg.u)
1927	: [uCmp] "r" (u8Old)
1928	, [uNew] "r" (u8New)
1929	, "[fXchg]" (0)
1930	RTASM_ARM_DMB_SY_COMMA_IN_REG
1931	: "cc");
1932	*pu8Old = u8ActualOld;
1933	# endif
1934	return fXchg.f;
1935
1936	# else
1937	# error "Port me"
1938	# endif
1939	}
1940	#endif
1941
1942
1943	/**
1944	* Atomically Compare and Exchange a signed 8-bit value, additionally
1945	* passes back old value, ordered.
1946	*
1947	* @returns true if xchg was done.
1948	* @returns false if xchg wasn't done.
1949	*
1950	* @param pi8 Pointer to the value to update.
1951	* @param i8New The new value to assigned to *pi8.
1952	* @param i8Old The old value to *pi8 compare with.
1953	* @param pi8Old Pointer store the old value at.
1954	*
1955	* @remarks x86: Requires a 486 or later.
1956	*/
1957	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
1958	{
1959	return ASMAtomicCmpXchgExU8((volatile uint8_t RT_FAR )pi8, (uint8_t)i8New, (uint8_t)i8Old, (uint8_t RT_FAR )pi8Old);
1960	}
1961
1962
1963	/**
1964	* Atomically Compare and Exchange an unsigned 16-bit value, additionally passes
1965	* back old value, ordered.
1966	*
1967	* @returns true if xchg was done.
1968	* @returns false if xchg wasn't done.
1969	*
1970	* @param pu16 Pointer to the value to update.
1971	* @param u16New The new value to assigned to *pu16.
1972	* @param u16Old The old value to *pu32 compare with.
1973	* @param pu16Old Pointer store the old value at.
1974	*
1975	* @remarks x86: Requires a 486 or later.
1976	*/
1977	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
1978	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;
1979	#else
1980	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
1981	{
1982	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1983	# if RT_INLINE_ASM_GNU_STYLE
1984	uint8_t u8Ret;
1985	__asm__ __volatile__("lock; cmpxchgw %3, %0\n\t"
1986	"setz %1\n\t"
1987	: "=m" (*pu16)
1988	, "=qm" (u8Ret)
1989	, "=a" (*pu16Old)
1990	: "r" (u16New)
1991	, "a" (u16Old)
1992	, "m" (*pu16)
1993	: "cc");
1994	return (bool)u8Ret;
1995
1996	# elif RT_INLINE_ASM_USES_INTRIN
1997	return (pu16Old = _InterlockedCompareExchange16((short RT_FAR )pu16, u16New, u16Old)) == u16Old;
1998
1999	# else
2000	uint16_t u16Ret;
2001	__asm
2002	{
2003	# ifdef RT_ARCH_AMD64
2004	mov rdx, [pu16]
2005	# else
2006	mov edx, [pu16]
2007	# endif
2008	mov eax, [u16Old]
2009	mov ecx, [u16New]
2010	# ifdef RT_ARCH_AMD64
2011	lock cmpxchg [rdx], ecx
2012	mov rdx, [pu16Old]
2013	mov [rdx], eax
2014	# else
2015	lock cmpxchg [edx], ecx
2016	mov edx, [pu16Old]
2017	mov [edx], eax
2018	# endif
2019	setz al
2020	movzx eax, al
2021	mov [u16Ret], eax
2022	}
2023	return !!u16Ret;
2024	# endif
2025
2026	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2027	/* M1 bench: match: casalh= 6577 vs dmb+cash= 1608 vs non-lse=5078 (ps/call)
2028	mismatch: casalh=18791 vs dmb+cash=19721 vs non-lse=2543 (ps/call) */
2029	# if defined(RTASM_ARM64_USE_FEAT_LSE)
2030	union { uint32_t u; bool f; } fXchg;
2031	uint32_t u32Actual;
2032	__asm__ __volatile__("Lstart_ASMAtomicCmpXchgExU16_%=:\n\t"
2033	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
2034	"casalh %w[uOldActual], %w[uNew], %[pMem]\n\t"
2035	# else
2036	RTASM_ARM_DMB_SY
2037	"cash %w[uOldActual], %w[uNew], %[pMem]\n\t"
2038	# endif
2039	"cmp %w[uOldActual], %w[uOldOrg]\n\t"
2040	"cset %w[fXchg], eq\n\t"
2041	: [pMem] "+Q" (*pu16)
2042	, [uOldActual] "=&r" (u32Actual)
2043	, [fXchg] "=&r" (fXchg.u)
2044	: [uNew] "r" ((uint32_t)u16New)
2045	, [uOldOrg] "r" ((uint32_t)u16Old)
2046	, "[uOldActual]" ((uint32_t)u16Old)
2047	: "cc");
2048	*pu16Old = (uint16_t)u32Actual;
2049	# else
2050	union { uint16_t u; bool f; } fXchg;
2051	uint16_t u16ActualOld;
2052	uint16_t rcSpill;
2053	__asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgExU16_%=:\n\t"
2054	RTASM_ARM_DMB_SY
2055	# if defined(RT_ARCH_ARM64)
2056	"ldaxrh %w[uOld], %[pMem]\n\t"
2057	"cmp %w[uOld], %w[uCmp]\n\t"
2058	"bne 1f\n\t" /* stop here if not equal */
2059	"stlxrh %w[rc], %w[uNew], %[pMem]\n\t"
2060	"cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgExU16_%=\n\t"
2061	"mov %w[fXchg], #1\n\t"
2062	"1:\n\t"
2063	"clrex\n\t"
2064	# else
2065	"ldrexh %[uOld], %[pMem]\n\t"
2066	"teq %[uOld], %[uCmp]\n\t"
2067	"strexheq %[rc], %[uNew], %[pMem]\n\t"
2068	"bne 1f\n\t" /* stop here if not equal */
2069	"cmp %[rc], #0\n\t"
2070	"bne Ltry_again_ASMAtomicCmpXchgExU16_%=\n\t"
2071	"mov %[fXchg], #1\n\t"
2072	"1:\n\t"
2073	/** @todo clrexne on armv7? */
2074	# endif
2075	: [pMem] "+Q" (*pu16)
2076	, [uOld] "=&r" (u16ActualOld)
2077	, [rc] "=&r" (rcSpill)
2078	, [fXchg] "=&r" (fXchg.u)
2079	: [uCmp] "r" (u16Old)
2080	, [uNew] "r" (u16New)
2081	, "[fXchg]" (0)
2082	RTASM_ARM_DMB_SY_COMMA_IN_REG
2083	: "cc");
2084	*pu16Old = u16ActualOld;
2085	# endif
2086	return fXchg.f;
2087
2088	# else
2089	# error "Port me"
2090	# endif
2091	}
2092	#endif
2093
2094
2095	/**
2096	* Atomically Compare and Exchange a signed 16-bit value, additionally
2097	* passes back old value, ordered.
2098	*
2099	* @returns true if xchg was done.
2100	* @returns false if xchg wasn't done.
2101	*
2102	* @param pi16 Pointer to the value to update.
2103	* @param i16New The new value to assigned to *pi16.
2104	* @param i16Old The old value to *pi16 compare with.
2105	* @param pi16Old Pointer store the old value at.
2106	*
2107	* @remarks x86: Requires a 486 or later.
2108	*/
2109	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
2110	{
2111	return ASMAtomicCmpXchgExU16((volatile uint16_t RT_FAR )pi16, (uint16_t)i16New, (uint16_t)i16Old, (uint16_t RT_FAR )pi16Old);
2112	}
2113
2114
2115	/**
2116	* Atomically Compare and Exchange an unsigned 32-bit value, additionally
2117	* passes back old value, ordered.
2118	*
2119	* @returns true if xchg was done.
2120	* @returns false if xchg wasn't done.
2121	*
2122	* @param pu32 Pointer to the value to update.
2123	* @param u32New The new value to assigned to *pu32.
2124	* @param u32Old The old value to *pu32 compare with.
2125	* @param pu32Old Pointer store the old value at.
2126	*
2127	* @remarks x86: Requires a 486 or later.
2128	*/
2129	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
2130	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;
2131	#else
2132	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
2133	{
2134	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
2135	# if RT_INLINE_ASM_GNU_STYLE
2136	uint8_t u8Ret;
2137	__asm__ __volatile__("lock; cmpxchgl %3, %0\n\t"
2138	"setz %1\n\t"
2139	: "=m" (*pu32)
2140	, "=qm" (u8Ret)
2141	, "=a" (*pu32Old)
2142	: "r" (u32New)
2143	, "a" (u32Old)
2144	, "m" (*pu32)
2145	: "cc");
2146	return (bool)u8Ret;
2147
2148	# elif RT_INLINE_ASM_USES_INTRIN
2149	return (pu32Old = _InterlockedCompareExchange((long RT_FAR )pu32, u32New, u32Old)) == u32Old;
2150
2151	# else
2152	uint32_t u32Ret;
2153	__asm
2154	{
2155	# ifdef RT_ARCH_AMD64
2156	mov rdx, [pu32]
2157	# else
2158	mov edx, [pu32]
2159	# endif
2160	mov eax, [u32Old]
2161	mov ecx, [u32New]
2162	# ifdef RT_ARCH_AMD64
2163	lock cmpxchg [rdx], ecx
2164	mov rdx, [pu32Old]
2165	mov [rdx], eax
2166	# else
2167	lock cmpxchg [edx], ecx
2168	mov edx, [pu32Old]
2169	mov [edx], eax
2170	# endif
2171	setz al
2172	movzx eax, al
2173	mov [u32Ret], eax
2174	}
2175	return !!u32Ret;
2176	# endif
2177
2178	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2179	union { uint32_t u; bool f; } fXchg;
2180	/* M1 bench: match: casal= 6590 vs dmb+cas= 1564 vs non-lse=5033 (ps/call)
2181	mismatch: casal=18790 vs dmb+cas=19711 vs non-lse=2503 (ps/call) */
2182	# if defined(RTASM_ARM64_USE_FEAT_LSE)
2183	__asm__ __volatile__("Lstart_ASMAtomicCmpXchgExU32_%=:\n\t"
2184	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
2185	"casal %w[uOldActual], %w[uNew], %[pMem]\n\t"
2186	# else
2187	RTASM_ARM_DMB_SY
2188	"cas %w[uOldActual], %w[uNew], %[pMem]\n\t"
2189	# endif
2190	"cmp %w[uOldActual], %w[uOldOrg]\n\t"
2191	"cset %w[fXchg], eq\n\t"
2192	: [pMem] "+Q" (*pu32)
2193	, [uOldActual] "=&r" (*pu32Old)
2194	, [fXchg] "=&r" (fXchg.u)
2195	: [uNew] "r" (u32New)
2196	, [uOldOrg] "r" (u32Old)
2197	, "[uOldActual]" (u32Old)
2198	: "cc");
2199	# else
2200	uint32_t u32ActualOld;
2201	uint32_t rcSpill;
2202	__asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgExU32_%=:\n\t"
2203	RTASM_ARM_DMB_SY
2204	# if defined(RT_ARCH_ARM64)
2205	"ldaxr %w[uOld], %[pMem]\n\t"
2206	"cmp %w[uOld], %w[uCmp]\n\t"
2207	"bne 1f\n\t" /* stop here if not equal */
2208	"stlxr %w[rc], %w[uNew], %[pMem]\n\t"
2209	"cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgExU32_%=\n\t"
2210	"mov %w[fXchg], #1\n\t"
2211	"1:\n\t"
2212	"clrex\n\t"
2213	# else
2214	"ldrex %[uOld], %[pMem]\n\t"
2215	"teq %[uOld], %[uCmp]\n\t"
2216	"strexeq %[rc], %[uNew], %[pMem]\n\t"
2217	"bne 1f\n\t" /* stop here if not equal */
2218	"cmp %[rc], #0\n\t"
2219	"bne Ltry_again_ASMAtomicCmpXchgExU32_%=\n\t"
2220	"mov %[fXchg], #1\n\t"
2221	"1:\n\t"
2222	/** @todo clrexne on armv7? */
2223	# endif
2224	: [pMem] "+Q" (*pu32)
2225	, [uOld] "=&r" (u32ActualOld)
2226	, [rc] "=&r" (rcSpill)
2227	, [fXchg] "=&r" (fXchg.u)
2228	: [uCmp] "r" (u32Old)
2229	, [uNew] "r" (u32New)
2230	, "[fXchg]" (0)
2231	RTASM_ARM_DMB_SY_COMMA_IN_REG
2232	: "cc");
2233	*pu32Old = u32ActualOld;
2234	# endif
2235	return fXchg.f;
2236
2237	# else
2238	# error "Port me"
2239	# endif
2240	}
2241	#endif
2242
2243
2244	/**
2245	* Atomically Compare and Exchange a signed 32-bit value, additionally
2246	* passes back old value, ordered.
2247	*
2248	* @returns true if xchg was done.
2249	* @returns false if xchg wasn't done.
2250	*
2251	* @param pi32 Pointer to the value to update.
2252	* @param i32New The new value to assigned to *pi32.
2253	* @param i32Old The old value to *pi32 compare with.
2254	* @param pi32Old Pointer store the old value at.
2255	*
2256	* @remarks x86: Requires a 486 or later.
2257	*/
2258	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
2259	{
2260	return ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR )pi32, (uint32_t)i32New, (uint32_t)i32Old, (uint32_t RT_FAR )pi32Old);
2261	}
2262
2263
2264	/**
2265	* Atomically Compare and exchange an unsigned 64-bit value, additionally
2266	* passing back old value, ordered.
2267	*
2268	* @returns true if xchg was done.
2269	* @returns false if xchg wasn't done.
2270	*
2271	* @param pu64 Pointer to the 64-bit variable to update.
2272	* @param u64New The 64-bit value to assign to *pu64.
2273	* @param u64Old The value to compare with.
2274	* @param pu64Old Pointer store the old value at.
2275	*
2276	* @remarks x86: Requires a Pentium or later.
2277	*/
2278	#if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \
2279	\|\| RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
2280	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;
2281	#else
2282	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
2283	{
2284	# if RT_INLINE_ASM_USES_INTRIN
2285	return (pu64Old =_InterlockedCompareExchange64((__int64 RT_FAR )pu64, u64New, u64Old)) == u64Old;
2286
2287	# elif defined(RT_ARCH_AMD64)
2288	# if RT_INLINE_ASM_GNU_STYLE
2289	uint8_t u8Ret;
2290	__asm__ __volatile__("lock; cmpxchgq %3, %0\n\t"
2291	"setz %1\n\t"
2292	: "=m" (*pu64)
2293	, "=qm" (u8Ret)
2294	, "=a" (*pu64Old)
2295	: "r" (u64New)
2296	, "a" (u64Old)
2297	, "m" (*pu64)
2298	: "cc");
2299	return (bool)u8Ret;
2300	# else
2301	bool fRet;
2302	__asm
2303	{
2304	mov rdx, [pu32]
2305	mov rax, [u64Old]
2306	mov rcx, [u64New]
2307	lock cmpxchg [rdx], rcx
2308	mov rdx, [pu64Old]
2309	mov [rdx], rax
2310	setz al
2311	mov [fRet], al
2312	}
2313	return fRet;
2314	# endif
2315
2316	# elif defined(RT_ARCH_X86)
2317	# if RT_INLINE_ASM_GNU_STYLE
2318	uint64_t u64Ret;
2319	# if defined(PIC) \|\| defined(__PIC__)
2320	/* Note #1: This code uses a memory clobber description, because the clean
2321	solution with an output value for *pu64 makes gcc run out of
2322	registers. This will cause suboptimal code, and anyone with a
2323	better solution is welcome to improve this.
2324
2325	Note #2: We must prevent gcc from encoding the memory access, as it
2326	may go via the GOT if we're working on a global variable (like
2327	in the testcase). Thus we request a register (%3) and
2328	dereference it ourselves. */
2329	__asm__ __volatile__("xchgl %%ebx, %1\n\t"
2330	"lock; cmpxchg8b (%3)\n\t"
2331	"xchgl %%ebx, %1\n\t"
2332	: "=A" (u64Ret)
2333	: "DS" ((uint32_t)u64New)
2334	, "c" ((uint32_t)(u64New >> 32))
2335	, "r" (pu64) /* Do not use "m" here*/
2336	, "0" (u64Old)
2337	: "memory"
2338	, "cc" );
2339	# else /* !PIC */
2340	__asm__ __volatile__("lock; cmpxchg8b %4\n\t"
2341	: "=A" (u64Ret)
2342	, "=m" (*pu64)
2343	: "b" ((uint32_t)u64New)
2344	, "c" ((uint32_t)(u64New >> 32))
2345	, "m" (*pu64)
2346	, "0" (u64Old)
2347	: "cc");
2348	# endif
2349	*pu64Old = u64Ret;
2350	return u64Ret == u64Old;
2351	# else
2352	uint32_t u32Ret;
2353	__asm
2354	{
2355	mov ebx, dword ptr [u64New]
2356	mov ecx, dword ptr [u64New + 4]
2357	mov edi, [pu64]
2358	mov eax, dword ptr [u64Old]
2359	mov edx, dword ptr [u64Old + 4]
2360	lock cmpxchg8b [edi]
2361	mov ebx, [pu64Old]
2362	mov [ebx], eax
2363	setz al
2364	movzx eax, al
2365	add ebx, 4
2366	mov [ebx], edx
2367	mov dword ptr [u32Ret], eax
2368	}
2369	return !!u32Ret;
2370	# endif
2371
2372	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2373	union { uint32_t u; bool f; } fXchg;
2374	/* M1 bench: match: casal= 6606 vs dmb+cas= 1565 vs non-lse=5006 (ps/call)
2375	mismatch: casal=18786 vs dmb+cas=19718 vs non-lse=2503 (ps/call) */
2376	# if defined(RTASM_ARM64_USE_FEAT_LSE)
2377	__asm__ __volatile__("Lstart_ASMAtomicCmpXchgExU32_%=:\n\t"
2378	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
2379	"casal %[uOldActual], %[uNew], %[pMem]\n\t"
2380	# else
2381	RTASM_ARM_DMB_SY
2382	"cas %[uOldActual], %[uNew], %[pMem]\n\t"
2383	# endif
2384	"cmp %[uOldActual], %[uOldOrg]\n\t"
2385	"cset %w[fXchg], eq\n\t"
2386	: [pMem] "+Q" (*pu64)
2387	, [uOldActual] "=&r" (*pu64Old)
2388	, [fXchg] "=&r" (fXchg.u)
2389	: [uNew] "r" (u64New)
2390	, [uOldOrg] "r" (u64Old)
2391	, "[uOldActual]" (u64Old)
2392	: "cc");
2393	# else
2394	uint64_t u64ActualOld;
2395	uint32_t rcSpill;
2396	__asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgU64_%=:\n\t"
2397	RTASM_ARM_DMB_SY
2398	# if defined(RT_ARCH_ARM64)
2399	"ldaxr %[uOld], %[pMem]\n\t"
2400	"cmp %[uOld], %[uCmp]\n\t"
2401	"bne 1f\n\t" /* stop here if not equal */
2402	"stlxr %w[rc], %[uNew], %[pMem]\n\t"
2403	"cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
2404	"mov %w[fXchg], #1\n\t"
2405	"1:\n\t"
2406	"clrex\n\t"
2407	# else
2408	"ldrexd %[uOld], %H[uOld], %[pMem]\n\t"
2409	"teq %[uOld], %[uCmp]\n\t"
2410	"teqeq %H[uOld], %H[uCmp]\n\t"
2411	"strexdeq %[rc], %[uNew], %H[uNew], %[pMem]\n\t"
2412	"bne 1f\n\t" /* stop here if not equal */
2413	"cmp %[rc], #0\n\t"
2414	"bne Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
2415	"mov %[fXchg], #1\n\t"
2416	"1:\n\t"
2417	/** @todo clrexne on armv7? */
2418	# endif
2419	: [pMem] "+Q" (*pu64)
2420	, [uOld] "=&r" (u64ActualOld)
2421	, [rc] "=&r" (rcSpill)
2422	, [fXchg] "=&r" (fXchg.u)
2423	: [uCmp] "r" (u64Old)
2424	, [uNew] "r" (u64New)
2425	, "[fXchg]" (0)
2426	RTASM_ARM_DMB_SY_COMMA_IN_REG
2427	: "cc");
2428	*pu64Old = u64ActualOld;
2429	# endif
2430	return fXchg.f;
2431
2432	# else
2433	# error "Port me"
2434	# endif
2435	}
2436	#endif
2437
2438
2439	/**
2440	* Atomically Compare and exchange a signed 64-bit value, additionally
2441	* passing back old value, ordered.
2442	*
2443	* @returns true if xchg was done.
2444	* @returns false if xchg wasn't done.
2445	*
2446	* @param pi64 Pointer to the 64-bit variable to update.
2447	* @param i64 The 64-bit value to assign to *pu64.
2448	* @param i64Old The value to compare with.
2449	* @param pi64Old Pointer store the old value at.
2450	*
2451	* @remarks x86: Requires a Pentium or later.
2452	*/
2453	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
2454	{
2455	return ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR )pi64, (uint64_t)i64, (uint64_t)i64Old, (uint64_t RT_FAR )pi64Old);
2456	}
2457
2458	#if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_ARM64) \|\| defined(DOXYGEN_RUNNING)
2459
2460	/** @def RTASM_HAVE_CMP_XCHG_U128
2461	* Indicates that we've got ASMAtomicCmpSwapU128(), ASMAtomicCmpSwapU128v2()
2462	* and ASMAtomicCmpSwapExU128() available. */
2463	# define RTASM_HAVE_CMP_XCHG_U128 1
2464
2465
2466	/**
2467	* Atomically compare and exchange an unsigned 128-bit value, ordered.
2468	*
2469	* @returns true if exchange was done.
2470	* @returns false if exchange wasn't done.
2471	*
2472	* @param pu128 Pointer to the 128-bit variable to update.
2473	* @param u64NewHi The high 64 bits of the value to assign to *pu128.
2474	* @param u64NewLo The low 64 bits of the value to assign to *pu128.
2475	* @param u64OldHi The high 64-bit of the value to compare with.
2476	* @param u64OldLo The low 64-bit of the value to compare with.
2477	* @param pu128Old Where to return the old value.
2478	*
2479	* @remarks AMD64: Not present in the earliest CPUs, so check CPUID.
2480	*/
2481	# if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN)
2482	DECLASM(bool) ASMAtomicCmpXchgU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo,
2483	const uint64_t u64OldHi, const uint64_t u64OldLo, uint128_t *pu128Old) RT_NOTHROW_PROTO;
2484	# else
2485	DECLINLINE(bool) ASMAtomicCmpXchgU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo,
2486	const uint64_t u64OldHi, const uint64_t u64OldLo, uint128_t *pu128Old) RT_NOTHROW_DEF
2487	{
2488	# if RT_INLINE_ASM_USES_INTRIN
2489	pu128Old->Hi = u64OldHi;
2490	pu128Old->Lo = u64OldLo;
2491	AssertCompileMemberOffset(uint128_t, Lo, 0);
2492	return _InterlockedCompareExchange128((__int64 volatile )pu128, u64NewHi, u64NewLo, (__int64 )&pu128Old->Lo) != 0;
2493
2494	# elif (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
2495	uint128_t const uCmp = ((uint128_t)u64OldHi << 64) \| u64OldLo;
2496	uint128_t const uOld = __sync_val_compare_and_swap(pu128, uCmp, ((uint128_t)u64NewHi << 64) \| u64NewLo);
2497	*pu128Old = uOld;
2498	return uCmp == uOld;
2499
2500	# elif defined(RT_ARCH_AMD64)
2501	# if RT_INLINE_ASM_GNU_STYLE
2502	uint8_t bRet;
2503	uint64_t u64RetHi, u64RetLo;
2504	__asm__ __volatile__("lock; cmpxchg16b %3\n\t"
2505	"setz %b0\n\t"
2506	: "=r" (bRet)
2507	, "=a" (u64RetLo)
2508	, "=d" (u64RetHi)
2509	, "+m" (*pu128)
2510	: "a" (u64OldLo)
2511	, "d" (u64OldHi)
2512	, "b" (u64NewLo)
2513	, "c" (u64NewHi)
2514	: "cc");
2515	*pu128Old = ((uint128_t)u64RetHi << 64) \| u64RetLo;
2516	return (bool)bRet;
2517	# else
2518	# error "Port me"
2519	# endif
2520	# else
2521	# error "Port me"
2522	# endif
2523	}
2524	# endif
2525
2526
2527	/**
2528	* Atomically compare and exchange an unsigned 128-bit value, ordered.
2529	*
2530	* @returns true if exchange was done.
2531	* @returns false if exchange wasn't done.
2532	*
2533	* @param pu128 Pointer to the 128-bit variable to update.
2534	* @param u128New The 128-bit value to assign to *pu128.
2535	* @param u128Old The value to compare with.
2536	* @param pu128Old Where to return the old value.
2537	*
2538	* @remarks AMD64: Not present in the earliest CPUs, so check CPUID.
2539	*/
2540	DECLINLINE(bool) ASMAtomicCmpXchgU128(volatile uint128_t *pu128, const uint128_t u128New,
2541	const uint128_t u128Old, uint128_t *pu128Old) RT_NOTHROW_DEF
2542	{
2543	# ifdef RT_COMPILER_WITH_128BIT_INT_TYPES
2544	# if (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
2545	uint128_t const uSwapped = __sync_val_compare_and_swap(pu128, u128Old, u128New);
2546	*pu128Old = uSwapped;
2547	return uSwapped == u128Old;
2548	# else
2549	return ASMAtomicCmpXchgU128v2(pu128, (uint64_t)(u128New >> 64), (uint64_t)u128New,
2550	(uint64_t)(u128Old >> 64), (uint64_t)u128Old, pu128Old);
2551	# endif
2552	# else
2553	return ASMAtomicCmpXchgU128v2(pu128, u128New.Hi, u128New.Lo, u128Old.Hi, u128Old.Lo, pu128Old);
2554	# endif
2555	}
2556
2557
2558	/**
2559	* RTUINT128U wrapper for ASMAtomicCmpXchgU128.
2560	*/
2561	DECLINLINE(bool) ASMAtomicCmpXchgU128U(volatile RTUINT128U *pu128, const RTUINT128U u128New,
2562	const RTUINT128U u128Old, PRTUINT128U pu128Old) RT_NOTHROW_DEF
2563	{
2564	# if (defined(__clang_major__) \|\| defined(__GNUC__)) && defined(RT_ARCH_ARM64)
2565	return ASMAtomicCmpXchgU128(&pu128->u, u128New.u, u128Old.u, &pu128Old->u);
2566	# else
2567	return ASMAtomicCmpXchgU128v2(&pu128->u, u128New.s.Hi, u128New.s.Lo, u128Old.s.Hi, u128Old.s.Lo, &pu128Old->u);
2568	# endif
2569	}
2570
2571	#endif /* RT_ARCH_AMD64 \|\| RT_ARCH_ARM64 */
2572
2573
2574
2575	/** @def ASMAtomicCmpXchgExHandle
2576	* Atomically Compare and Exchange a typical IPRT handle value, ordered.
2577	*
2578	* @param ph Pointer to the value to update.
2579	* @param hNew The new value to assigned to *pu.
2580	* @param hOld The old value to *pu compare with.
2581	* @param fRc Where to store the result.
2582	* @param phOldVal Pointer to where to store the old value.
2583	*
2584	* @remarks This doesn't currently work for all handles (like RTFILE).
2585	*/
2586	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
2587	# define ASMAtomicCmpXchgExHandle(ph, hNew, hOld, fRc, phOldVal) \
2588	do { \
2589	AssertCompile(sizeof(*ph) == sizeof(uint32_t)); \
2590	AssertCompile(sizeof(*phOldVal) == sizeof(uint32_t)); \
2591	(fRc) = ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR )(ph), (uint32_t)(hNew), (uint32_t)(hOld), (uint32_t RT_FAR )(phOldVal)); \
2592	} while (0)
2593	#elif HC_ARCH_BITS == 64
2594	# define ASMAtomicCmpXchgExHandle(ph, hNew, hOld, fRc, phOldVal) \
2595	do { \
2596	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
2597	AssertCompile(sizeof(*(phOldVal)) == sizeof(uint64_t)); \
2598	(fRc) = ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR )(ph), (uint64_t)(hNew), (uint64_t)(hOld), (uint64_t RT_FAR )(phOldVal)); \
2599	} while (0)
2600	#else
2601	# error HC_ARCH_BITS
2602	#endif
2603
2604
2605	/** @def ASMAtomicCmpXchgExSize
2606	* Atomically Compare and Exchange a value which size might differ
2607	* between platforms or compilers. Additionally passes back old value.
2608	*
2609	* @param pu Pointer to the value to update.
2610	* @param uNew The new value to assigned to *pu.
2611	* @param uOld The old value to *pu compare with.
2612	* @param fRc Where to store the result.
2613	* @param puOldVal Pointer to where to store the old value.
2614	*
2615	* @remarks x86: Requires a 486 or later.
2616	*/
2617	#define ASMAtomicCmpXchgExSize(pu, uNew, uOld, fRc, puOldVal) \
2618	do { \
2619	switch (sizeof(*(pu))) { \
2620	case 4: (fRc) = ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew), (uint32_t)(uOld), (uint32_t RT_FAR *)(uOldVal)); \
2621	break; \
2622	case 8: (fRc) = ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew), (uint64_t)(uOld), (uint64_t RT_FAR *)(uOldVal)); \
2623	break; \
2624	default: AssertMsgFailed(("ASMAtomicCmpXchgSize: size %d is not supported\n", sizeof(*(pu)))); \
2625	(fRc) = false; \
2626	(uOldVal) = 0; \
2627	break; \
2628	} \
2629	} while (0)
2630
2631
2632	/**
2633	* Atomically Compare and Exchange a pointer value, additionally
2634	* passing back old value, ordered.
2635	*
2636	* @returns true if xchg was done.
2637	* @returns false if xchg wasn't done.
2638	*
2639	* @param ppv Pointer to the value to update.
2640	* @param pvNew The new value to assigned to *ppv.
2641	* @param pvOld The old value to *ppv compare with.
2642	* @param ppvOld Pointer store the old value at.
2643	*
2644	* @remarks x86: Requires a 486 or later.
2645	*/
2646	DECLINLINE(bool) ASMAtomicCmpXchgExPtrVoid(void RT_FAR * volatile RT_FAR ppv, const void RT_FAR pvNew, const void RT_FAR *pvOld,
2647	void RT_FAR * RT_FAR *ppvOld) RT_NOTHROW_DEF
2648	{
2649	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
2650	return ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR )(void RT_FAR )ppv, (uint32_t)pvNew, (uint32_t)pvOld, (uint32_t RT_FAR *)ppvOld);
2651	#elif ARCH_BITS == 64
2652	return ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR )(void RT_FAR )ppv, (uint64_t)pvNew, (uint64_t)pvOld, (uint64_t RT_FAR *)ppvOld);
2653	#else
2654	# error "ARCH_BITS is bogus"
2655	#endif
2656	}
2657
2658
2659	/**
2660	* Atomically Compare and Exchange a pointer value, additionally
2661	* passing back old value, ordered.
2662	*
2663	* @returns true if xchg was done.
2664	* @returns false if xchg wasn't done.
2665	*
2666	* @param ppv Pointer to the value to update.
2667	* @param pvNew The new value to assigned to *ppv.
2668	* @param pvOld The old value to *ppv compare with.
2669	* @param ppvOld Pointer store the old value at.
2670	*
2671	* @remarks This is relatively type safe on GCC platforms.
2672	* @remarks x86: Requires a 486 or later.
2673	*/
2674	#ifdef __GNUC__
2675	# define ASMAtomicCmpXchgExPtr(ppv, pvNew, pvOld, ppvOld) \
2676	__extension__ \
2677	({\
2678	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
2679	__typeof__(*(ppv)) const pvNewTypeChecked = (pvNew); \
2680	__typeof__(*(ppv)) const pvOldTypeChecked = (pvOld); \
2681	__typeof__((ppv)) const ppvOldTypeChecked = (ppvOld); \
2682	bool fMacroRet = ASMAtomicCmpXchgExPtrVoid((void * volatile *)ppvTypeChecked, \
2683	(void )pvNewTypeChecked, (void )pvOldTypeChecked, \
2684	(void **)ppvOldTypeChecked); \
2685	fMacroRet; \
2686	})
2687	#else
2688	# define ASMAtomicCmpXchgExPtr(ppv, pvNew, pvOld, ppvOld) \
2689	ASMAtomicCmpXchgExPtrVoid((void RT_FAR * volatile RT_FAR )(ppv), (void RT_FAR )(pvNew), (void RT_FAR )(pvOld), (void RT_FAR RT_FAR *)(ppvOld))
2690	#endif
2691
2692
2693	/**
2694	* Virtualization unfriendly serializing instruction, always exits.
2695	*/
2696	#if (RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN) \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
2697	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSerializeInstructionCpuId(void) RT_NOTHROW_PROTO;
2698	#else
2699	DECLINLINE(void) ASMSerializeInstructionCpuId(void) RT_NOTHROW_DEF
2700	{
2701	# if RT_INLINE_ASM_GNU_STYLE
2702	RTCCUINTREG xAX = 0;
2703	# ifdef RT_ARCH_AMD64
2704	__asm__ __volatile__ ("cpuid"
2705	: "=a" (xAX)
2706	: "0" (xAX)
2707	: "rbx", "rcx", "rdx", "memory");
2708	# elif (defined(PIC) \|\| defined(__PIC__)) && defined(__i386__)
2709	__asm__ __volatile__ ("push %%ebx\n\t"
2710	"cpuid\n\t"
2711	"pop %%ebx\n\t"
2712	: "=a" (xAX)
2713	: "0" (xAX)
2714	: "ecx", "edx", "memory");
2715	# else
2716	__asm__ __volatile__ ("cpuid"
2717	: "=a" (xAX)
2718	: "0" (xAX)
2719	: "ebx", "ecx", "edx", "memory");
2720	# endif
2721
2722	# elif RT_INLINE_ASM_USES_INTRIN
2723	int aInfo[4];
2724	_ReadWriteBarrier();
2725	__cpuid(aInfo, 0);
2726
2727	# else
2728	__asm
2729	{
2730	push ebx
2731	xor eax, eax
2732	cpuid
2733	pop ebx
2734	}
2735	# endif
2736	}
2737	#endif
2738
2739	/**
2740	* Virtualization friendly serializing instruction, though more expensive.
2741	*/
2742	#if RT_INLINE_ASM_EXTERNAL \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
2743	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSerializeInstructionIRet(void) RT_NOTHROW_PROTO;
2744	#else
2745	DECLINLINE(void) ASMSerializeInstructionIRet(void) RT_NOTHROW_DEF
2746	{
2747	# if RT_INLINE_ASM_GNU_STYLE
2748	# ifdef RT_ARCH_AMD64
2749	__asm__ __volatile__ ("movq %%rsp,%%r10\n\t"
2750	"subq $128, %%rsp\n\t" /redzone/
2751	"mov %%ss, %%eax\n\t"
2752	"pushq %%rax\n\t"
2753	"pushq %%r10\n\t"
2754	"pushfq\n\t"
2755	"movl %%cs, %%eax\n\t"
2756	"pushq %%rax\n\t"
2757	"leaq 1f(%%rip), %%rax\n\t"
2758	"pushq %%rax\n\t"
2759	"iretq\n\t"
2760	"1:\n\t"
2761	::: "rax", "r10", "memory", "cc");
2762	# else
2763	__asm__ __volatile__ ("pushfl\n\t"
2764	"pushl %%cs\n\t"
2765	"pushl $1f\n\t"
2766	"iretl\n\t"
2767	"1:\n\t"
2768	::: "memory");
2769	# endif
2770
2771	# else
2772	__asm
2773	{
2774	pushfd
2775	push cs
2776	push la_ret
2777	iretd
2778	la_ret:
2779	}
2780	# endif
2781	}
2782	#endif
2783
2784	/**
2785	* Virtualization friendlier serializing instruction, may still cause exits.
2786	*/
2787	#if (RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2008) \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
2788	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSerializeInstructionRdTscp(void) RT_NOTHROW_PROTO;
2789	#else
2790	DECLINLINE(void) ASMSerializeInstructionRdTscp(void) RT_NOTHROW_DEF
2791	{
2792	# if RT_INLINE_ASM_GNU_STYLE
2793	/* rdtscp is not supported by ancient linux build VM of course :-( */
2794	# ifdef RT_ARCH_AMD64
2795	/__asm__ __volatile__("rdtscp\n\t" ::: "rax", "rdx, "rcx"); /
2796	__asm__ __volatile__(".byte 0x0f,0x01,0xf9\n\t" ::: "rax", "rdx", "rcx", "memory");
2797	# else
2798	/__asm__ __volatile__("rdtscp\n\t" ::: "eax", "edx, "ecx"); /
2799	__asm__ __volatile__(".byte 0x0f,0x01,0xf9\n\t" ::: "eax", "edx", "ecx", "memory");
2800	# endif
2801	# else
2802	# if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2008
2803	uint32_t uIgnore;
2804	_ReadWriteBarrier();
2805	(void)__rdtscp(&uIgnore);
2806	(void)uIgnore;
2807	# else
2808	__asm
2809	{
2810	rdtscp
2811	}
2812	# endif
2813	# endif
2814	}
2815	#endif
2816
2817
2818	/**
2819	* Serialize Instruction (both data store and instruction flush).
2820	*/
2821	#if (defined(RT_ARCH_X86) && ARCH_BITS == 16) \|\| defined(IN_GUEST)
2822	# define ASMSerializeInstruction() ASMSerializeInstructionIRet()
2823	#elif defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)
2824	# define ASMSerializeInstruction() ASMSerializeInstructionCpuId()
2825	#elif defined(RT_ARCH_SPARC64)
2826	RTDECL(void) ASMSerializeInstruction(void) RT_NOTHROW_PROTO;
2827	#elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2828	DECLINLINE(void) ASMSerializeInstruction(void) RT_NOTHROW_DEF
2829	{
2830	__asm__ __volatile__ (RTASM_ARM_DSB_SY :: RTASM_ARM_DSB_SY_IN_REG :);
2831	}
2832	#else
2833	# error "Port me"
2834	#endif
2835
2836
2837	/**
2838	* Memory fence, waits for any pending writes and reads to complete.
2839	* @note No implicit compiler barrier (which is probably stupid).
2840	*/
2841	DECLINLINE(void) ASMMemoryFence(void) RT_NOTHROW_DEF
2842	{
2843	#if defined(RT_ARCH_AMD64) \|\| (defined(RT_ARCH_X86) && !defined(RT_WITH_OLD_CPU_SUPPORT))
2844	# if RT_INLINE_ASM_GNU_STYLE
2845	__asm__ __volatile__ (".byte 0x0f,0xae,0xf0\n\t");
2846	# elif RT_INLINE_ASM_USES_INTRIN
2847	_mm_mfence();
2848	# else
2849	__asm
2850	{
2851	_emit 0x0f
2852	_emit 0xae
2853	_emit 0xf0
2854	}
2855	# endif
2856	#elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2857	__asm__ __volatile__ (RTASM_ARM_DMB_SY :: RTASM_ARM_DMB_SY_IN_REG :);
2858	#elif ARCH_BITS == 16
2859	uint16_t volatile u16;
2860	ASMAtomicXchgU16(&u16, 0);
2861	#else
2862	uint32_t volatile u32;
2863	ASMAtomicXchgU32(&u32, 0);
2864	#endif
2865	}
2866
2867
2868	/**
2869	* Write fence, waits for any pending writes to complete.
2870	* @note No implicit compiler barrier (which is probably stupid).
2871	*/
2872	DECLINLINE(void) ASMWriteFence(void) RT_NOTHROW_DEF
2873	{
2874	#if defined(RT_ARCH_AMD64) \|\| (defined(RT_ARCH_X86) && !defined(RT_WITH_OLD_CPU_SUPPORT))
2875	# if RT_INLINE_ASM_GNU_STYLE
2876	__asm__ __volatile__ (".byte 0x0f,0xae,0xf8\n\t");
2877	# elif RT_INLINE_ASM_USES_INTRIN
2878	_mm_sfence();
2879	# else
2880	__asm
2881	{
2882	_emit 0x0f
2883	_emit 0xae
2884	_emit 0xf8
2885	}
2886	# endif
2887	#elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2888	__asm__ __volatile__ (RTASM_ARM_DMB_ST :: RTASM_ARM_DMB_ST_IN_REG :);
2889	#else
2890	ASMMemoryFence();
2891	#endif
2892	}
2893
2894
2895	/**
2896	* Read fence, waits for any pending reads to complete.
2897	* @note No implicit compiler barrier (which is probably stupid).
2898	*/
2899	DECLINLINE(void) ASMReadFence(void) RT_NOTHROW_DEF
2900	{
2901	#if defined(RT_ARCH_AMD64) \|\| (defined(RT_ARCH_X86) && !defined(RT_WITH_OLD_CPU_SUPPORT))
2902	# if RT_INLINE_ASM_GNU_STYLE
2903	__asm__ __volatile__ (".byte 0x0f,0xae,0xe8\n\t");
2904	# elif RT_INLINE_ASM_USES_INTRIN
2905	_mm_lfence();
2906	# else
2907	__asm
2908	{
2909	_emit 0x0f
2910	_emit 0xae
2911	_emit 0xe8
2912	}
2913	# endif
2914	#elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2915	__asm__ __volatile__ (RTASM_ARM_DMB_LD :: RTASM_ARM_DMB_LD_IN_REG :);
2916	#else
2917	ASMMemoryFence();
2918	#endif
2919	}
2920
2921
2922	/**
2923	* Atomically reads an unsigned 8-bit value, ordered.
2924	*
2925	* @returns Current *pu8 value
2926	* @param pu8 Pointer to the 8-bit variable to read.
2927	*/
2928	DECLINLINE(uint8_t) ASMAtomicReadU8(volatile uint8_t RT_FAR *pu8) RT_NOTHROW_DEF
2929	{
2930	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2931	uint32_t u32;
2932	# if defined(RTASM_ARM64_USE_FEAT_LSE) && 0 /* very expensive on M1 */
2933	__asm__ __volatile__("Lstart_ASMAtomicReadU8_%=:\n\t"
2934	RTASM_ARM_DMB_SY
2935	"casab %w[uDst], wzr, %[pMem]\n\t"
2936	: [uDst] "=&r" (u32)
2937	: [pMem] "Q" (*pu8),
2938	"0" (0)
2939	RTASM_ARM_DMB_SY_COMMA_IN_REG);
2940	# else
2941	__asm__ __volatile__("Lstart_ASMAtomicReadU8_%=:\n\t"
2942	RTASM_ARM_DMB_SY
2943	# if defined(RT_ARCH_ARM64)
2944	# if 1 /* shouldn't be any need for more than single-copy atomicity when we've got a proper barrier, just like on x86. */
2945	"ldurb %w[uDst], %[pMem]\n\t"
2946	# else
2947	"ldxrb %w[uDst], %[pMem]\n\t"
2948	"clrex\n\t"
2949	# endif
2950	# else
2951	"ldrexb %[uDst], %[pMem]\n\t"
2952	/** @todo clrex */
2953	# endif
2954	: [uDst] "=&r" (u32)
2955	: [pMem] "Q" (*pu8)
2956	RTASM_ARM_DMB_SY_COMMA_IN_REG);
2957	# endif
2958	return (uint8_t)u32;
2959	#else
2960	ASMMemoryFence();
2961	return pu8; / byte reads are atomic on x86 */
2962	#endif
2963	}
2964
2965
2966	/**
2967	* Atomically reads an unsigned 8-bit value, unordered.
2968	*
2969	* @returns Current *pu8 value
2970	* @param pu8 Pointer to the 8-bit variable to read.
2971	*/
2972	DECLINLINE(uint8_t) ASMAtomicUoReadU8(volatile uint8_t RT_FAR *pu8) RT_NOTHROW_DEF
2973	{
2974	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
2975	uint32_t u32;
2976	__asm__ __volatile__("Lstart_ASMAtomicUoReadU8_%=:\n\t"
2977	# if defined(RT_ARCH_ARM64)
2978	"ldurb %w[uDst], %[pMem]\n\t"
2979	# else
2980	"ldrexb %[uDst], %[pMem]\n\t" /** @todo fix this */
2981	# endif
2982	: [uDst] "=&r" (u32)
2983	: [pMem] "Q" (*pu8));
2984	return (uint8_t)u32;
2985	#else
2986	return pu8; / byte reads are atomic on x86 */
2987	#endif
2988	}
2989
2990
2991	/**
2992	* Atomically reads a signed 8-bit value, ordered.
2993	*
2994	* @returns Current *pi8 value
2995	* @param pi8 Pointer to the 8-bit variable to read.
2996	*/
2997	DECLINLINE(int8_t) ASMAtomicReadS8(volatile int8_t RT_FAR *pi8) RT_NOTHROW_DEF
2998	{
2999	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3000	return (int8_t)ASMAtomicReadU8((volatile uint8_t RT_FAR *)pi8);
3001	#else
3002	ASMMemoryFence();
3003	return pi8; / byte reads are atomic on x86 */
3004	#endif
3005	}
3006
3007
3008	/**
3009	* Atomically reads a signed 8-bit value, unordered.
3010	*
3011	* @returns Current *pi8 value
3012	* @param pi8 Pointer to the 8-bit variable to read.
3013	*/
3014	DECLINLINE(int8_t) ASMAtomicUoReadS8(volatile int8_t RT_FAR *pi8) RT_NOTHROW_DEF
3015	{
3016	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3017	int32_t i32;
3018	__asm__ __volatile__("Lstart_ASMAtomicUoReadS8_%=:\n\t"
3019	# if defined(RT_ARCH_ARM64)
3020	"ldurb %w[iDst], %[pMem]\n\t"
3021	# else
3022	"ldrexb %[iDst], %[pMem]\n\t" /** @todo fix this */
3023	# endif
3024	: [iDst] "=&r" (i32)
3025	: [pMem] "Q" (*pi8));
3026	return (int8_t)i32;
3027	#else
3028	return pi8; / byte reads are atomic on x86 */
3029	#endif
3030	}
3031
3032
3033	/**
3034	* Atomically reads an unsigned 16-bit value, ordered.
3035	*
3036	* @returns Current *pu16 value
3037	* @param pu16 Pointer to the 16-bit variable to read.
3038	*/
3039	DECLINLINE(uint16_t) ASMAtomicReadU16(volatile uint16_t RT_FAR *pu16) RT_NOTHROW_DEF
3040	{
3041	Assert(!((uintptr_t)pu16 & 1));
3042	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3043	uint32_t u32;
3044	# if defined(RTASM_ARM64_USE_FEAT_LSE) && 0 /* very expensive on M1, but alignment advantages with LEA2 (M2?). */
3045	__asm__ __volatile__("Lstart_ASMAtomicReadU16_%=:\n\t"
3046	RTASM_ARM_DMB_SY
3047	"casah %w[uDst], wzr, %[pMem]\n\t"
3048	: [uDst] "=&r" (u32)
3049	: [pMem] "Q" (*pu16),
3050	"0" (0)
3051	RTASM_ARM_DMB_SY_COMMA_IN_REG);
3052	# else
3053	__asm__ __volatile__("Lstart_ASMAtomicReadU16_%=:\n\t"
3054	RTASM_ARM_DMB_SY
3055	# if defined(RT_ARCH_ARM64)
3056	# if 1 /* ASSUMING proper barrier and aligned access, we should be fine with single-copy atomicity, just like on x86. */
3057	"ldurh %w[uDst], %[pMem]\n\t"
3058	# else
3059	"ldxrh %w[uDst], %[pMem]\n\t"
3060	"clrex\n\t"
3061	# endif
3062	# else
3063	"ldrexh %[uDst], %[pMem]\n\t"
3064	/** @todo clrex */
3065	# endif
3066	: [uDst] "=&r" (u32)
3067	: [pMem] "Q" (*pu16)
3068	RTASM_ARM_DMB_SY_COMMA_IN_REG);
3069	# endif
3070	return (uint16_t)u32;
3071	#else
3072	ASMMemoryFence();
3073	return *pu16;
3074	#endif
3075	}
3076
3077
3078	/**
3079	* Atomically reads an unsigned 16-bit value, unordered.
3080	*
3081	* @returns Current *pu16 value
3082	* @param pu16 Pointer to the 16-bit variable to read.
3083	*/
3084	DECLINLINE(uint16_t) ASMAtomicUoReadU16(volatile uint16_t RT_FAR *pu16) RT_NOTHROW_DEF
3085	{
3086	Assert(!((uintptr_t)pu16 & 1));
3087	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3088	uint32_t u32;
3089	__asm__ __volatile__("Lstart_ASMAtomicUoReadU16_%=:\n\t"
3090	# if defined(RT_ARCH_ARM64)
3091	"ldurh %w[uDst], %[pMem]\n\t"
3092	# else
3093	"ldrexh %[uDst], %[pMem]\n\t" /** @todo fix this */
3094	# endif
3095	: [uDst] "=&r" (u32)
3096	: [pMem] "Q" (*pu16));
3097	return (uint16_t)u32;
3098	#else
3099	return *pu16;
3100	#endif
3101	}
3102
3103
3104	/**
3105	* Atomically reads a signed 16-bit value, ordered.
3106	*
3107	* @returns Current *pi16 value
3108	* @param pi16 Pointer to the 16-bit variable to read.
3109	*/
3110	DECLINLINE(int16_t) ASMAtomicReadS16(volatile int16_t RT_FAR *pi16) RT_NOTHROW_DEF
3111	{
3112	Assert(!((uintptr_t)pi16 & 1));
3113	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3114	return (int16_t)ASMAtomicReadU16((volatile uint16_t RT_FAR *)pi16);
3115	#else
3116	ASMMemoryFence();
3117	return *pi16;
3118	#endif
3119	}
3120
3121
3122	/**
3123	* Atomically reads a signed 16-bit value, unordered.
3124	*
3125	* @returns Current *pi16 value
3126	* @param pi16 Pointer to the 16-bit variable to read.
3127	*/
3128	DECLINLINE(int16_t) ASMAtomicUoReadS16(volatile int16_t RT_FAR *pi16) RT_NOTHROW_DEF
3129	{
3130	Assert(!((uintptr_t)pi16 & 1));
3131	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3132	int32_t i32;
3133	__asm__ __volatile__("Lstart_ASMAtomicUoReadS16_%=:\n\t"
3134	# if defined(RT_ARCH_ARM64)
3135	"ldurh %w[iDst], %[pMem]\n\t"
3136	# else
3137	"ldrexh %[iDst], %[pMem]\n\t" /** @todo fix this */
3138	# endif
3139	: [iDst] "=&r" (i32)
3140	: [pMem] "Q" (*pi16));
3141	return (int16_t)i32;
3142	#else
3143	return *pi16;
3144	#endif
3145	}
3146
3147
3148	/**
3149	* Atomically reads an unsigned 32-bit value, ordered.
3150	*
3151	* @returns Current *pu32 value
3152	* @param pu32 Pointer to the 32-bit variable to read.
3153	*/
3154	DECLINLINE(uint32_t) ASMAtomicReadU32(volatile uint32_t RT_FAR *pu32) RT_NOTHROW_DEF
3155	{
3156	Assert(!((uintptr_t)pu32 & 3));
3157	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3158	uint32_t u32;
3159	# if defined(RTASM_ARM64_USE_FEAT_LSE) && 0 /* very expensive on M1, but alignment advantages with LEA2 (M2?). */
3160	__asm__ __volatile__("Lstart_ASMAtomicReadU32_%=:\n\t"
3161	RTASM_ARM_DMB_SY
3162	"casa %w[uDst], wzr, %[pMem]\n\t"
3163	: [uDst] "=&r" (u32)
3164	: [pMem] "Q" (*pu32),
3165	"0" (0)
3166	RTASM_ARM_DMB_SY_COMMA_IN_REG);
3167	# else
3168	__asm__ __volatile__("Lstart_ASMAtomicReadU32_%=:\n\t"
3169	RTASM_ARM_DMB_SY
3170	# if defined(RT_ARCH_ARM64)
3171	# if 1 /* ASSUMING proper barrier and aligned access, we should be fine with single-copy atomicity, just like on x86. */
3172	"ldur %w[uDst], %[pMem]\n\t"
3173	# else
3174	"ldxr %w[uDst], %[pMem]\n\t"
3175	"clrex\n\t"
3176	# endif
3177	# else
3178	"ldrex %[uDst], %[pMem]\n\t"
3179	/** @todo clrex */
3180	# endif
3181	: [uDst] "=&r" (u32)
3182	: [pMem] "Q" (*pu32)
3183	RTASM_ARM_DMB_SY_COMMA_IN_REG);
3184	# endif
3185	return u32;
3186	#else
3187	ASMMemoryFence();
3188	# if ARCH_BITS == 16
3189	AssertFailed(); /** @todo 16-bit */
3190	# endif
3191	return *pu32;
3192	#endif
3193	}
3194
3195
3196	/**
3197	* Atomically reads an unsigned 32-bit value, unordered.
3198	*
3199	* @returns Current *pu32 value
3200	* @param pu32 Pointer to the 32-bit variable to read.
3201	*/
3202	DECLINLINE(uint32_t) ASMAtomicUoReadU32(volatile uint32_t RT_FAR *pu32) RT_NOTHROW_DEF
3203	{
3204	Assert(!((uintptr_t)pu32 & 3));
3205	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3206	uint32_t u32;
3207	__asm__ __volatile__("Lstart_ASMAtomicUoReadU32_%=:\n\t"
3208	# if defined(RT_ARCH_ARM64)
3209	"ldur %w[uDst], %[pMem]\n\t"
3210	# else
3211	"ldrex %[uDst], %[pMem]\n\t" /** @todo fix this */
3212	# endif
3213	: [uDst] "=&r" (u32)
3214	: [pMem] "Q" (*pu32));
3215	return u32;
3216	#else
3217	# if ARCH_BITS == 16
3218	AssertFailed(); /** @todo 16-bit */
3219	# endif
3220	return *pu32;
3221	#endif
3222	}
3223
3224
3225	/**
3226	* Atomically reads a signed 32-bit value, ordered.
3227	*
3228	* @returns Current *pi32 value
3229	* @param pi32 Pointer to the 32-bit variable to read.
3230	*/
3231	DECLINLINE(int32_t) ASMAtomicReadS32(volatile int32_t RT_FAR *pi32) RT_NOTHROW_DEF
3232	{
3233	Assert(!((uintptr_t)pi32 & 3));
3234	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3235	return (int32_t)ASMAtomicReadU32((volatile uint32_t RT_FAR *)pi32);
3236	#else
3237	ASMMemoryFence();
3238	# if ARCH_BITS == 16
3239	AssertFailed(); /** @todo 16-bit */
3240	# endif
3241	return *pi32;
3242	#endif
3243	}
3244
3245
3246	/**
3247	* Atomically reads a signed 32-bit value, unordered.
3248	*
3249	* @returns Current *pi32 value
3250	* @param pi32 Pointer to the 32-bit variable to read.
3251	*/
3252	DECLINLINE(int32_t) ASMAtomicUoReadS32(volatile int32_t RT_FAR *pi32) RT_NOTHROW_DEF
3253	{
3254	Assert(!((uintptr_t)pi32 & 3));
3255	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3256	int32_t i32;
3257	__asm__ __volatile__("Lstart_ASMAtomicUoReadS32_%=:\n\t"
3258	# if defined(RT_ARCH_ARM64)
3259	"ldur %w[iDst], %[pMem]\n\t"
3260	# else
3261	"ldrex %[iDst], %[pMem]\n\t" /** @todo thix this */
3262	# endif
3263	: [iDst] "=&r" (i32)
3264	: [pMem] "Q" (*pi32));
3265	return i32;
3266
3267	#else
3268	# if ARCH_BITS == 16
3269	AssertFailed(); /** @todo 16-bit */
3270	# endif
3271	return *pi32;
3272	#endif
3273	}
3274
3275
3276	/**
3277	* Atomically reads an unsigned 64-bit value, ordered.
3278	*
3279	* @returns Current *pu64 value
3280	* @param pu64 Pointer to the 64-bit variable to read.
3281	* The memory pointed to must be writable.
3282	*
3283	* @remarks This may fault if the memory is read-only!
3284	* @remarks x86: Requires a Pentium or later.
3285	*/
3286	#if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !defined(RT_ARCH_AMD64)) \
3287	\|\| RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC
3288	RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_PROTO;
3289	#else
3290	DECLINLINE(uint64_t) ASMAtomicReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_DEF
3291	{
3292	uint64_t u64;
3293	# ifdef RT_ARCH_AMD64
3294	Assert(!((uintptr_t)pu64 & 7));
3295	/*# if RT_INLINE_ASM_GNU_STYLE
3296	__asm__ __volatile__( "mfence\n\t"
3297	"movq %1, %0\n\t"
3298	: "=r" (u64)
3299	: "m" (*pu64));
3300	# else
3301	__asm
3302	{
3303	mfence
3304	mov rdx, [pu64]
3305	mov rax, [rdx]
3306	mov [u64], rax
3307	}
3308	# endif*/
3309	ASMMemoryFence();
3310	u64 = *pu64;
3311
3312	# elif defined(RT_ARCH_X86)
3313	# if RT_INLINE_ASM_GNU_STYLE
3314	# if defined(PIC) \|\| defined(__PIC__)
3315	uint32_t u32EBX = 0;
3316	Assert(!((uintptr_t)pu64 & 7));
3317	__asm__ __volatile__("xchgl %%ebx, %3\n\t"
3318	"lock; cmpxchg8b (%5)\n\t"
3319	"movl %3, %%ebx\n\t"
3320	: "=A" (u64)
3321	# if RT_GNUC_PREREQ(4, 3)
3322	, "+m" (*pu64)
3323	# else
3324	, "=m" (*pu64)
3325	# endif
3326	: "0" (0ULL)
3327	, "m" (u32EBX)
3328	, "c" (0)
3329	, "S" (pu64)
3330	: "cc");
3331	# else /* !PIC */
3332	__asm__ __volatile__("lock; cmpxchg8b %1\n\t"
3333	: "=A" (u64)
3334	, "+m" (*pu64)
3335	: "0" (0ULL)
3336	, "b" (0)
3337	, "c" (0)
3338	: "cc");
3339	# endif
3340	# else
3341	Assert(!((uintptr_t)pu64 & 7));
3342	__asm
3343	{
3344	xor eax, eax
3345	xor edx, edx
3346	mov edi, pu64
3347	xor ecx, ecx
3348	xor ebx, ebx
3349	lock cmpxchg8b [edi]
3350	mov dword ptr [u64], eax
3351	mov dword ptr [u64 + 4], edx
3352	}
3353	# endif
3354
3355	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3356	Assert(!((uintptr_t)pu64 & 7));
3357
3358	# if defined(RTASM_ARM64_USE_FEAT_LSE) && 0 /* very expensive on M1, but alignment advantages with LEA2 (M2?). */
3359	__asm__ __volatile__("Lstart_ASMAtomicReadU64_%=:\n\t"
3360	RTASM_ARM_DMB_SY
3361	"casa %[uDst], xzr, %[pMem]\n\t"
3362	: [uDst] "=&r" (u64)
3363	: [pMem] "Q" (*pu64),
3364	"0" (0)
3365	RTASM_ARM_DMB_SY_COMMA_IN_REG);
3366	# else
3367	__asm__ __volatile__("Lstart_ASMAtomicReadU64_%=:\n\t"
3368	RTASM_ARM_DMB_SY
3369	# if defined(RT_ARCH_ARM64)
3370	# if 1 /* ASSUMING proper barrier and aligned access, we should be fine with single-copy atomicity, just like on x86. */
3371	"ldur %[uDst], %[pMem]\n\t"
3372	# else
3373	"ldxr %[uDst], %[pMem]\n\t"
3374	"clrex\n\t"
3375	# endif
3376	# else
3377	"ldrexd %[uDst], %H[uDst], %[pMem]\n\t"
3378	/** @todo clrex */
3379	# endif
3380	: [uDst] "=&r" (u64)
3381	: [pMem] "Q" (*pu64)
3382	RTASM_ARM_DMB_SY_COMMA_IN_REG);
3383	# endif
3384	# else
3385	# error "Port me"
3386	# endif
3387	return u64;
3388	}
3389	#endif
3390
3391
3392	/**
3393	* Atomically reads an unsigned 64-bit value, unordered.
3394	*
3395	* @returns Current *pu64 value
3396	* @param pu64 Pointer to the 64-bit variable to read.
3397	* The memory pointed to must be writable.
3398	*
3399	* @remarks This may fault if the memory is read-only!
3400	* @remarks x86: Requires a Pentium or later.
3401	*/
3402	#if !defined(RT_ARCH_AMD64) \
3403	&& ( (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \
3404	\|\| RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC)
3405	RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicUoReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_PROTO;
3406	#else
3407	DECLINLINE(uint64_t) ASMAtomicUoReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_DEF
3408	{
3409	uint64_t u64;
3410	# ifdef RT_ARCH_AMD64
3411	Assert(!((uintptr_t)pu64 & 7));
3412	/*# if RT_INLINE_ASM_GNU_STYLE
3413	Assert(!((uintptr_t)pu64 & 7));
3414	__asm__ __volatile__("movq %1, %0\n\t"
3415	: "=r" (u64)
3416	: "m" (*pu64));
3417	# else
3418	__asm
3419	{
3420	mov rdx, [pu64]
3421	mov rax, [rdx]
3422	mov [u64], rax
3423	}
3424	# endif */
3425	u64 = *pu64;
3426
3427	# elif defined(RT_ARCH_X86)
3428	# if RT_INLINE_ASM_GNU_STYLE
3429	# if defined(PIC) \|\| defined(__PIC__)
3430	uint32_t u32EBX = 0;
3431	uint32_t u32Spill;
3432	Assert(!((uintptr_t)pu64 & 7));
3433	__asm__ __volatile__("xor %%eax,%%eax\n\t"
3434	"xor %%ecx,%%ecx\n\t"
3435	"xor %%edx,%%edx\n\t"
3436	"xchgl %%ebx, %3\n\t"
3437	"lock; cmpxchg8b (%4)\n\t"
3438	"movl %3, %%ebx\n\t"
3439	: "=A" (u64)
3440	# if RT_GNUC_PREREQ(4, 3)
3441	, "+m" (*pu64)
3442	# else
3443	, "=m" (*pu64)
3444	# endif
3445	, "=c" (u32Spill)
3446	: "m" (u32EBX)
3447	, "S" (pu64)
3448	: "cc");
3449	# else /* !PIC */
3450	__asm__ __volatile__("lock; cmpxchg8b %1\n\t"
3451	: "=A" (u64)
3452	, "+m" (*pu64)
3453	: "0" (0ULL)
3454	, "b" (0)
3455	, "c" (0)
3456	: "cc");
3457	# endif
3458	# else
3459	Assert(!((uintptr_t)pu64 & 7));
3460	__asm
3461	{
3462	xor eax, eax
3463	xor edx, edx
3464	mov edi, pu64
3465	xor ecx, ecx
3466	xor ebx, ebx
3467	lock cmpxchg8b [edi]
3468	mov dword ptr [u64], eax
3469	mov dword ptr [u64 + 4], edx
3470	}
3471	# endif
3472
3473	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
3474	Assert(!((uintptr_t)pu64 & 7));
3475	__asm__ __volatile__("Lstart_ASMAtomicUoReadU64_%=:\n\t"
3476	# if defined(RT_ARCH_ARM64)
3477	"ldur %[uDst], %[pMem]\n\t"
3478	# else
3479	"ldrexd %[uDst], %H[uDst], %[pMem]\n\t" /* this is required for atomic access since it's a pair */
3480	/** @todo clrex? */
3481	# endif
3482	: [uDst] "=&r" (u64)
3483	: [pMem] "Q" (*pu64));
3484
3485	# else
3486	# error "Port me"
3487	# endif
3488	return u64;
3489	}
3490	#endif
3491
3492
3493	/**
3494	* Atomically reads a signed 64-bit value, ordered.
3495	*
3496	* @returns Current *pi64 value
3497	* @param pi64 Pointer to the 64-bit variable to read.
3498	* The memory pointed to must be writable.
3499	*
3500	* @remarks This may fault if the memory is read-only!
3501	* @remarks x86: Requires a Pentium or later.
3502	*/
3503	DECLINLINE(int64_t) ASMAtomicReadS64(volatile int64_t RT_FAR *pi64) RT_NOTHROW_DEF
3504	{
3505	return (int64_t)ASMAtomicReadU64((volatile uint64_t RT_FAR *)pi64);
3506	}
3507
3508
3509	/**
3510	* Atomically reads a signed 64-bit value, unordered.
3511	*
3512	* @returns Current *pi64 value
3513	* @param pi64 Pointer to the 64-bit variable to read.
3514	* The memory pointed to must be writable.
3515	*
3516	* @remarks This will fault if the memory is read-only!
3517	* @remarks x86: Requires a Pentium or later.
3518	*/
3519	DECLINLINE(int64_t) ASMAtomicUoReadS64(volatile int64_t RT_FAR *pi64) RT_NOTHROW_DEF
3520	{
3521	return (int64_t)ASMAtomicUoReadU64((volatile uint64_t RT_FAR *)pi64);
3522	}
3523
3524
3525	/** @def RTASM_HAVE_READ_U128
3526	* Defined in the target architecture supports atomic reading of 128-bit
3527	* integers.
3528	*
3529	* The define value is zero if both ordered and unordered reads are implemented
3530	* using ASMAtomicCmpXchgU128v2(). It is 1 if unordered reads are done natively
3531	* w/o cmpxchg and 3 if both variants are done natively w/o cmpxchg.
3532	*
3533	* @note AMD64: Caller must check for cmpxchg16b support before use and make
3534	* sure variables are writable (won't be changed).
3535	* @sa RTASM_HAVE_CMP_XCHG_U128, RTASM_HAVE_WRITE_U128
3536	*/
3537	#if defined(RT_ARCH_ARM64) \|\| defined(DOXYGEN_RUNNING)
3538	# define RTASM_HAVE_READ_U128 3
3539	#elif defined(RTASM_HAVE_CMP_XCHG_U128)
3540	# define RTASM_HAVE_READ_U128 0
3541	#endif
3542
3543	#ifdef RTASM_HAVE_READ_U128
3544
3545	/**
3546	* Atomically reads an unsigned 128-bit value, ordered.
3547	*
3548	* @returns Current *pu128 value
3549	* @param pu128 Pointer to the 128-bit variable to read.
3550	* The memory pointed to must be writable.
3551	*
3552	* @remarks AMD64: Requires the memory to be both readable and writable.
3553	* @remarks AMD64: Requires support for cmpxchg16b.
3554	*/
3555	DECLINLINE(uint128_t) ASMAtomicReadU128(volatile uint128_t RT_FAR *pu128) RT_NOTHROW_DEF
3556	{
3557	RTUINT128U u128Ret;
3558	Assert(!((uintptr_t)pu128 & 15));
3559	# if defined(__GNUC__) && defined(RT_ARCH_ARM64)
3560	__asm__ __volatile__("Lstart_ASMAtomicReadU128_%=:\n\t"
3561	RTASM_ARM_DMB_SY
3562	"ldp %[uRetLo], %[uRetHi], %[pMem]\n\t"
3563	RTASM_ARM_DMB_SY
3564	: [uRetHi] "=r" (u128Ret.s.Hi)
3565	, [uRetLo] "=r" (u128Ret.s.Lo)
3566	: [pMem] "Q" (*pu128)
3567	: );
3568	return u128Ret.u;
3569	# else
3570	ASMAtomicCmpXchgU128v2(pu128, 0, 0, 0, 0, &u128Ret.u);
3571	return u128Ret.u;
3572	# endif
3573	}
3574
3575	/**
3576	* Atomically reads an unsigned 128-bit value, ordered.
3577	*
3578	* @returns Current *pu128 value
3579	* @param pu128 Pointer to the 128-bit variable to read.
3580	* The memory pointed to must be writable.
3581	*
3582	* @remarks AMD64: Requires the memory to be both readable and writable.
3583	* @remarks AMD64: Requires support for cmpxchg16b.
3584	*/
3585	DECLINLINE(RTUINT128U) ASMAtomicReadU128U(volatile RTUINT128U RT_FAR *pu128) RT_NOTHROW_DEF
3586	{
3587	RTUINT128U u128Ret;
3588	Assert(!((uintptr_t)pu128 & 15));
3589	# if defined(__GNUC__) && defined(RT_ARCH_ARM64)
3590	__asm__ __volatile__("Lstart_ASMAtomicReadU128U_%=:\n\t"
3591	RTASM_ARM_DMB_SY
3592	"ldp %[uRetLo], %[uRetHi], %[pMem]\n\t"
3593	RTASM_ARM_DMB_SY
3594	: [uRetHi] "=r" (u128Ret.s.Hi)
3595	, [uRetLo] "=r" (u128Ret.s.Lo)
3596	: [pMem] "Q" (*pu128)
3597	: );
3598	return u128Ret;
3599	# else
3600	ASMAtomicCmpXchgU128v2(&pu128->u, 0, 0, 0, 0, &u128Ret.u);
3601	return u128Ret;
3602	# endif
3603	}
3604
3605
3606	/**
3607	* Atomically reads an unsigned 128-bit value, unordered.
3608	*
3609	* @returns Current *pu128 value
3610	* @param pu128 Pointer to the 128-bit variable to read.
3611	* The memory pointed to must be writable.
3612	*
3613	* @remarks AMD64: Requires the memory to be both readable and writable.
3614	* @remarks AMD64: Requires support for cmpxchg16b.
3615	* @remarks AMD64: Is ordered.
3616	*/
3617	DECLINLINE(uint128_t) ASMAtomicUoReadU128(volatile uint128_t RT_FAR *pu128) RT_NOTHROW_DEF
3618	{
3619	Assert(!((uintptr_t)pu128 & 15));
3620	# if defined(__GNUC__) && defined(RT_ARCH_ARM64)
3621	RTUINT128U u128Ret;
3622	__asm__ __volatile__("Lstart_ASMAtomicUoReadU128_%=:\n\t"
3623	"ldp %[uRetLo], %[uRetHi], %[pMem]\n\t"
3624	: [uRetHi] "=r" (u128Ret.s.Hi)
3625	, [uRetLo] "=r" (u128Ret.s.Lo)
3626	: [pMem] "Q" (*pu128)
3627	: );
3628	return u128Ret.u;
3629
3630	# elif defined(RT_ARCH_AMD64) && 0
3631	/* This doesn't work because __m128i can't be made volatile and we're not
3632	able to force MSC (2019) to emit _mm_load_si128 (besides it emits movdqu
3633	instead of movdqa). */
3634	__m128i uTmpSse = _mm_load_si128((__m128i volatile *)pu128);
3635	__m128i uTmpSseHi = _mm_srli_si128(uTmpSse, 64 / 8);
3636	RTUINT128U u128Ret;
3637	u128Ret.s.Lo = (uint64_t)_mm_cvtsi128_si64(uTmpSse);
3638	u128Ret.s.Hi = (uint64_t)_mm_cvtsi128_si64(uTmpSseHi);
3639	return u128Ret.u;
3640
3641	# else
3642	return ASMAtomicReadU128(pu128);
3643	# endif
3644	}
3645
3646	/**
3647	* Atomically reads an unsigned 128-bit value, unordered.
3648	*
3649	* @returns Current *pu128 value
3650	* @param pu128 Pointer to the 128-bit variable to read.
3651	* The memory pointed to must be writable.
3652	*
3653	* @remarks AMD64: Requires the memory to be both readable and writable.
3654	* @remarks AMD64: Requires support for cmpxchg16b.
3655	* @remarks AMD64: Is ordered.
3656	*/
3657	DECLINLINE(RTUINT128U) ASMAtomicUoReadU128U(volatile RTUINT128U RT_FAR *pu128) RT_NOTHROW_DEF
3658	{
3659	Assert(!((uintptr_t)pu128 & 15));
3660	# if defined(__GNUC__) && defined(RT_ARCH_ARM64)
3661	RTUINT128U u128Ret;
3662	__asm__ __volatile__("Lstart_ASMAtomicUoReadU128U_%=:\n\t"
3663	"ldp %[uRetLo], %[uRetHi], %[pMem]\n\t"
3664	: [uRetHi] "=r" (u128Ret.s.Hi)
3665	, [uRetLo] "=r" (u128Ret.s.Lo)
3666	: [pMem] "Q" (*pu128)
3667	: );
3668	return u128Ret;
3669	# else
3670	return ASMAtomicReadU128U(pu128);
3671	# endif
3672	}
3673
3674	#endif /* RTASM_HAVE_READ_U128 */
3675
3676	/**
3677	* Atomically reads a size_t value, ordered.
3678	*
3679	* @returns Current *pcb value
3680	* @param pcb Pointer to the size_t variable to read.
3681	*/
3682	DECLINLINE(size_t) ASMAtomicReadZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF
3683	{
3684	#if ARCH_BITS == 64
3685	return ASMAtomicReadU64((uint64_t volatile RT_FAR *)pcb);
3686	#elif ARCH_BITS == 32
3687	return ASMAtomicReadU32((uint32_t volatile RT_FAR *)pcb);
3688	#elif ARCH_BITS == 16
3689	AssertCompileSize(size_t, 2);
3690	return ASMAtomicReadU16((uint16_t volatile RT_FAR *)pcb);
3691	#else
3692	# error "Unsupported ARCH_BITS value"
3693	#endif
3694	}
3695
3696
3697	/**
3698	* Atomically reads a size_t value, unordered.
3699	*
3700	* @returns Current *pcb value
3701	* @param pcb Pointer to the size_t variable to read.
3702	*/
3703	DECLINLINE(size_t) ASMAtomicUoReadZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF
3704	{
3705	#if ARCH_BITS == 64 \|\| ARCH_BITS == 16
3706	return ASMAtomicUoReadU64((uint64_t volatile RT_FAR *)pcb);
3707	#elif ARCH_BITS == 32
3708	return ASMAtomicUoReadU32((uint32_t volatile RT_FAR *)pcb);
3709	#elif ARCH_BITS == 16
3710	AssertCompileSize(size_t, 2);
3711	return ASMAtomicUoReadU16((uint16_t volatile RT_FAR *)pcb);
3712	#else
3713	# error "Unsupported ARCH_BITS value"
3714	#endif
3715	}
3716
3717
3718	/**
3719	* Atomically reads a pointer value, ordered.
3720	*
3721	* @returns Current *pv value
3722	* @param ppv Pointer to the pointer variable to read.
3723	*
3724	* @remarks Please use ASMAtomicReadPtrT, it provides better type safety and
3725	* requires less typing (no casts).
3726	*/
3727	DECLINLINE(void RT_FAR ) ASMAtomicReadPtr(void RT_FAR volatile RT_FAR *ppv) RT_NOTHROW_DEF
3728	{
3729	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
3730	return (void RT_FAR )ASMAtomicReadU32((volatile uint32_t RT_FAR )(void RT_FAR *)ppv);
3731	#elif ARCH_BITS == 64
3732	return (void RT_FAR )ASMAtomicReadU64((volatile uint64_t RT_FAR )(void RT_FAR *)ppv);
3733	#else
3734	# error "ARCH_BITS is bogus"
3735	#endif
3736	}
3737
3738	/**
3739	* Convenience macro for avoiding the annoying casting with ASMAtomicReadPtr.
3740	*
3741	* @returns Current *pv value
3742	* @param ppv Pointer to the pointer variable to read.
3743	* @param Type The type of *ppv, sans volatile.
3744	*/
3745	#ifdef __GNUC__ /* 8.2.0 requires -Wno-ignored-qualifiers */
3746	# define ASMAtomicReadPtrT(ppv, Type) \
3747	__extension__ \
3748	({\
3749	__typeof__((ppv)) volatile ppvTypeChecked = (ppv); \
3750	Type pvTypeChecked = (__typeof__((ppv))) ASMAtomicReadPtr((void volatile *)ppvTypeChecked); \
3751	pvTypeChecked; \
3752	})
3753	#else
3754	# define ASMAtomicReadPtrT(ppv, Type) \
3755	(Type)ASMAtomicReadPtr((void RT_FAR * volatile RT_FAR *)(ppv))
3756	#endif
3757
3758
3759	/**
3760	* Atomically reads a pointer value, unordered.
3761	*
3762	* @returns Current *pv value
3763	* @param ppv Pointer to the pointer variable to read.
3764	*
3765	* @remarks Please use ASMAtomicUoReadPtrT, it provides better type safety and
3766	* requires less typing (no casts).
3767	*/
3768	DECLINLINE(void RT_FAR ) ASMAtomicUoReadPtr(void RT_FAR volatile RT_FAR *ppv) RT_NOTHROW_DEF
3769	{
3770	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
3771	return (void RT_FAR )ASMAtomicUoReadU32((volatile uint32_t RT_FAR )(void RT_FAR *)ppv);
3772	#elif ARCH_BITS == 64
3773	return (void RT_FAR )ASMAtomicUoReadU64((volatile uint64_t RT_FAR )(void RT_FAR *)ppv);
3774	#else
3775	# error "ARCH_BITS is bogus"
3776	#endif
3777	}
3778
3779
3780	/**
3781	* Convenience macro for avoiding the annoying casting with ASMAtomicUoReadPtr.
3782	*
3783	* @returns Current *pv value
3784	* @param ppv Pointer to the pointer variable to read.
3785	* @param Type The type of *ppv, sans volatile.
3786	*/
3787	#ifdef __GNUC__ /* 8.2.0 requires -Wno-ignored-qualifiers */
3788	# define ASMAtomicUoReadPtrT(ppv, Type) \
3789	__extension__ \
3790	({\
3791	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
3792	Type pvTypeChecked = (__typeof__((ppv))) ASMAtomicUoReadPtr((void volatile *)ppvTypeChecked); \
3793	pvTypeChecked; \
3794	})
3795	#else
3796	# define ASMAtomicUoReadPtrT(ppv, Type) \
3797	(Type)ASMAtomicUoReadPtr((void RT_FAR * volatile RT_FAR *)(ppv))
3798	#endif
3799
3800
3801	/**
3802	* Atomically reads a boolean value, ordered.
3803	*
3804	* @returns Current *pf value
3805	* @param pf Pointer to the boolean variable to read.
3806	*/
3807	DECLINLINE(bool) ASMAtomicReadBool(volatile bool RT_FAR *pf) RT_NOTHROW_DEF
3808	{
3809	ASMMemoryFence();
3810	return pf; / byte reads are atomic on x86 */
3811	}
3812
3813
3814	/**
3815	* Atomically reads a boolean value, unordered.
3816	*
3817	* @returns Current *pf value
3818	* @param pf Pointer to the boolean variable to read.
3819	*/
3820	DECLINLINE(bool) ASMAtomicUoReadBool(volatile bool RT_FAR *pf) RT_NOTHROW_DEF
3821	{
3822	return pf; / byte reads are atomic on x86 */
3823	}
3824
3825
3826	/**
3827	* Atomically read a typical IPRT handle value, ordered.
3828	*
3829	* @param ph Pointer to the handle variable to read.
3830	* @param phRes Where to store the result.
3831	*
3832	* @remarks This doesn't currently work for all handles (like RTFILE).
3833	*/
3834	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
3835	# define ASMAtomicReadHandle(ph, phRes) \
3836	do { \
3837	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
3838	AssertCompile(sizeof(*(phRes)) == sizeof(uint32_t)); \
3839	(uint32_t RT_FAR )(phRes) = ASMAtomicReadU32((uint32_t volatile RT_FAR *)(ph)); \
3840	} while (0)
3841	#elif HC_ARCH_BITS == 64
3842	# define ASMAtomicReadHandle(ph, phRes) \
3843	do { \
3844	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
3845	AssertCompile(sizeof(*(phRes)) == sizeof(uint64_t)); \
3846	(uint64_t RT_FAR )(phRes) = ASMAtomicReadU64((uint64_t volatile RT_FAR *)(ph)); \
3847	} while (0)
3848	#else
3849	# error HC_ARCH_BITS
3850	#endif
3851
3852
3853	/**
3854	* Atomically read a typical IPRT handle value, unordered.
3855	*
3856	* @param ph Pointer to the handle variable to read.
3857	* @param phRes Where to store the result.
3858	*
3859	* @remarks This doesn't currently work for all handles (like RTFILE).
3860	*/
3861	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
3862	# define ASMAtomicUoReadHandle(ph, phRes) \
3863	do { \
3864	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
3865	AssertCompile(sizeof(*(phRes)) == sizeof(uint32_t)); \
3866	(uint32_t RT_FAR )(phRes) = ASMAtomicUoReadU32((uint32_t volatile RT_FAR *)(ph)); \
3867	} while (0)
3868	#elif HC_ARCH_BITS == 64
3869	# define ASMAtomicUoReadHandle(ph, phRes) \
3870	do { \
3871	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
3872	AssertCompile(sizeof(*(phRes)) == sizeof(uint64_t)); \
3873	(uint64_t RT_FAR )(phRes) = ASMAtomicUoReadU64((uint64_t volatile RT_FAR *)(ph)); \
3874	} while (0)
3875	#else
3876	# error HC_ARCH_BITS
3877	#endif
3878
3879
3880	/**
3881	* Atomically read a value which size might differ
3882	* between platforms or compilers, ordered.
3883	*
3884	* @param pu Pointer to the variable to read.
3885	* @param puRes Where to store the result.
3886	*/
3887	#define ASMAtomicReadSize(pu, puRes) \
3888	do { \
3889	switch (sizeof(*(pu))) { \
3890	case 1: (uint8_t RT_FAR )(puRes) = ASMAtomicReadU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu)); break; \
3891	case 2: (uint16_t RT_FAR )(puRes) = ASMAtomicReadU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu)); break; \
3892	case 4: (uint32_t RT_FAR )(puRes) = ASMAtomicReadU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu)); break; \
3893	case 8: (uint64_t RT_FAR )(puRes) = ASMAtomicReadU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu)); break; \
3894	default: AssertMsgFailed(("ASMAtomicReadSize: size %d is not supported\n", sizeof(*(pu)))); \
3895	} \
3896	} while (0)
3897
3898
3899	/**
3900	* Atomically read a value which size might differ
3901	* between platforms or compilers, unordered.
3902	*
3903	* @param pu Pointer to the variable to read.
3904	* @param puRes Where to store the result.
3905	*/
3906	#define ASMAtomicUoReadSize(pu, puRes) \
3907	do { \
3908	switch (sizeof(*(pu))) { \
3909	case 1: (uint8_t RT_FAR )(puRes) = ASMAtomicUoReadU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu)); break; \
3910	case 2: (uint16_t RT_FAR )(puRes) = ASMAtomicUoReadU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu)); break; \
3911	case 4: (uint32_t RT_FAR )(puRes) = ASMAtomicUoReadU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu)); break; \
3912	case 8: (uint64_t RT_FAR )(puRes) = ASMAtomicUoReadU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu)); break; \
3913	default: AssertMsgFailed(("ASMAtomicReadSize: size %d is not supported\n", sizeof(*(pu)))); \
3914	} \
3915	} while (0)
3916
3917
3918	/**
3919	* Atomically writes an unsigned 8-bit value, ordered.
3920	*
3921	* @param pu8 Pointer to the 8-bit variable.
3922	* @param u8 The 8-bit value to assign to *pu8.
3923	*/
3924	DECLINLINE(void) ASMAtomicWriteU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_DEF
3925	{
3926	#if defined(RT_ARCH_ARM64)
3927	/* The DMB SY will ensure ordering a la x86, the stlrb is probably overkill
3928	as all byte accesses are single-copy atomic, which I think suffices here. */
3929	__asm__ __volatile__("Lstart_ASMAtomicWriteU8_%=:\n\t"
3930	# if defined(RTASM_ARM64_USE_FEAT_LSE) && 0 /* this is a lot slower and has no alignment benefits with LSE2 */
3931	RTASM_ARM_DMB_SY
3932	"swpb %w[uValue], wzr, %[pMem]\n\t"
3933	# else
3934	RTASM_ARM_DMB_SY
3935	"stlrb %w[uValue], %[pMem]\n\t" /* single-copy atomic w/ release semantics. */
3936	# endif
3937	: [pMem] "+Q" (*pu8)
3938	: [uValue] "r" ((uint32_t)u8)
3939	: );
3940	#else
3941	ASMAtomicXchgU8(pu8, u8);
3942	#endif
3943	}
3944
3945
3946	/**
3947	* Atomically writes an unsigned 8-bit value, unordered.
3948	*
3949	* @param pu8 Pointer to the 8-bit variable.
3950	* @param u8 The 8-bit value to assign to *pu8.
3951	*/
3952	DECLINLINE(void) ASMAtomicUoWriteU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_DEF
3953	{
3954	pu8 = u8; / byte writes are atomic on x86 */
3955	}
3956
3957
3958	/**
3959	* Atomically writes a signed 8-bit value, ordered.
3960	*
3961	* @param pi8 Pointer to the 8-bit variable to read.
3962	* @param i8 The 8-bit value to assign to *pi8.
3963	*/
3964	DECLINLINE(void) ASMAtomicWriteS8(volatile int8_t RT_FAR *pi8, int8_t i8) RT_NOTHROW_DEF
3965	{
3966	#if defined(RT_ARCH_ARM64)
3967	ASMAtomicWriteU8((volatile uint8_t RT_FAR *)pi8, (uint8_t)i8);
3968	#else
3969	ASMAtomicXchgS8(pi8, i8);
3970	#endif
3971	}
3972
3973
3974	/**
3975	* Atomically writes a signed 8-bit value, unordered.
3976	*
3977	* @param pi8 Pointer to the 8-bit variable to write.
3978	* @param i8 The 8-bit value to assign to *pi8.
3979	*/
3980	DECLINLINE(void) ASMAtomicUoWriteS8(volatile int8_t RT_FAR *pi8, int8_t i8) RT_NOTHROW_DEF
3981	{
3982	pi8 = i8; / byte writes are atomic on x86 */
3983	}
3984
3985
3986	/**
3987	* Atomically writes an unsigned 16-bit value, ordered.
3988	*
3989	* @param pu16 Pointer to the 16-bit variable to write.
3990	* @param u16 The 16-bit value to assign to *pu16.
3991	*/
3992	DECLINLINE(void) ASMAtomicWriteU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_DEF
3993	{
3994	#if defined(RT_ARCH_ARM64)
3995	__asm__ __volatile__("Lstart_ASMAtomicWriteU16_%=:\n\t"
3996	# if defined(RTASM_ARM64_USE_FEAT_LSE) /* slower on M1, but benefits from relaxed LSE2 alignment requirements (M2?). */
3997	RTASM_ARM_DMB_SY
3998	"swph %w[uValue], wzr, %[pMem]\n\t"
3999	# else
4000	RTASM_ARM_DMB_SY
4001	"stlrh %w[uValue], %[pMem]\n\t" /* single-copy atomic w/ release semantics. */
4002	# endif
4003	: [pMem] "+Q" (*pu16)
4004	: [uValue] "r" ((uint32_t)u16)
4005	: );
4006	#else
4007	ASMAtomicXchgU16(pu16, u16);
4008	#endif
4009	}
4010
4011
4012	/**
4013	* Atomically writes an unsigned 16-bit value, unordered.
4014	*
4015	* @param pu16 Pointer to the 16-bit variable to write.
4016	* @param u16 The 16-bit value to assign to *pu16.
4017	*/
4018	DECLINLINE(void) ASMAtomicUoWriteU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_DEF
4019	{
4020	Assert(!((uintptr_t)pu16 & 1));
4021	*pu16 = u16;
4022	}
4023
4024
4025	/**
4026	* Atomically writes a signed 16-bit value, ordered.
4027	*
4028	* @param pi16 Pointer to the 16-bit variable to write.
4029	* @param i16 The 16-bit value to assign to *pi16.
4030	*/
4031	DECLINLINE(void) ASMAtomicWriteS16(volatile int16_t RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF
4032	{
4033	#if defined(RT_ARCH_ARM64)
4034	ASMAtomicWriteU16((volatile uint16_t RT_FAR *)pi16, (uint16_t)i16);
4035	#else
4036	ASMAtomicXchgS16(pi16, i16);
4037	#endif
4038	}
4039
4040
4041	/**
4042	* Atomically writes a signed 16-bit value, unordered.
4043	*
4044	* @param pi16 Pointer to the 16-bit variable to write.
4045	* @param i16 The 16-bit value to assign to *pi16.
4046	*/
4047	DECLINLINE(void) ASMAtomicUoWriteS16(volatile int16_t RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF
4048	{
4049	Assert(!((uintptr_t)pi16 & 1));
4050	*pi16 = i16;
4051	}
4052
4053
4054	/**
4055	* Atomically writes an unsigned 32-bit value, ordered.
4056	*
4057	* @param pu32 Pointer to the 32-bit variable to write.
4058	* @param u32 The 32-bit value to assign to *pu32.
4059	*/
4060	DECLINLINE(void) ASMAtomicWriteU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
4061	{
4062	#if defined(RT_ARCH_ARM64)
4063	__asm__ __volatile__("Lstart_ASMAtomicWriteU32_%=:\n\t"
4064	# if defined(RTASM_ARM64_USE_FEAT_LSE) /* slower on M1, but benefits from relaxed LSE2 alignment requirements (M2?). */
4065	RTASM_ARM_DMB_SY
4066	"swp %w[uValue], wzr, %[pMem]\n\t"
4067	# else
4068	RTASM_ARM_DMB_SY
4069	"stlr %w[uValue], %[pMem]\n\t" /* single-copy atomic w/ release semantics. */
4070	# endif
4071	: [pMem] "+Q" (*pu32)
4072	: [uValue] "r" (u32)
4073	: "cc");
4074	#else
4075	ASMAtomicXchgU32(pu32, u32);
4076	#endif
4077	}
4078
4079
4080	/**
4081	* Atomically writes an unsigned 32-bit value, unordered.
4082	*
4083	* @param pu32 Pointer to the 32-bit variable to write.
4084	* @param u32 The 32-bit value to assign to *pu32.
4085	*/
4086	DECLINLINE(void) ASMAtomicUoWriteU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
4087	{
4088	Assert(!((uintptr_t)pu32 & 3));
4089	#if ARCH_BITS >= 32
4090	*pu32 = u32;
4091	#else
4092	ASMAtomicXchgU32(pu32, u32);
4093	#endif
4094	}
4095
4096
4097	/**
4098	* Atomically writes a signed 32-bit value, ordered.
4099	*
4100	* @param pi32 Pointer to the 32-bit variable to write.
4101	* @param i32 The 32-bit value to assign to *pi32.
4102	*/
4103	DECLINLINE(void) ASMAtomicWriteS32(volatile int32_t RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
4104	{
4105	#if defined(RT_ARCH_ARM64)
4106	ASMAtomicWriteU32((volatile uint32_t RT_FAR *)pi32, (uint32_t)i32);
4107	#else
4108	ASMAtomicXchgS32(pi32, i32);
4109	#endif
4110	}
4111
4112
4113	/**
4114	* Atomically writes a signed 32-bit value, unordered.
4115	*
4116	* @param pi32 Pointer to the 32-bit variable to write.
4117	* @param i32 The 32-bit value to assign to *pi32.
4118	*/
4119	DECLINLINE(void) ASMAtomicUoWriteS32(volatile int32_t RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
4120	{
4121	Assert(!((uintptr_t)pi32 & 3));
4122	#if ARCH_BITS >= 32
4123	*pi32 = i32;
4124	#else
4125	ASMAtomicXchgS32(pi32, i32);
4126	#endif
4127	}
4128
4129
4130	/**
4131	* Atomically writes an unsigned 64-bit value, ordered.
4132	*
4133	* @param pu64 Pointer to the 64-bit variable to write.
4134	* @param u64 The 64-bit value to assign to *pu64.
4135	*/
4136	DECLINLINE(void) ASMAtomicWriteU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
4137	{
4138	#if defined(RT_ARCH_ARM64)
4139	__asm__ __volatile__("Lstart_ASMAtomicWriteU64_%=:\n\t"
4140	# if defined(RTASM_ARM64_USE_FEAT_LSE) /* slower on M1, but benefits from relaxed LSE2 alignment requirements (M2?). */
4141	RTASM_ARM_DMB_SY
4142	"swp %[uValue], xzr, %[pMem]\n\t"
4143	# else
4144	RTASM_ARM_DMB_SY /** @todo necessary? */
4145	"stlr %[uValue], %[pMem]\n\t"
4146	# endif
4147	: [pMem] "+Q" (*pu64)
4148	: [uValue] "r" (u64)
4149	: );
4150	#else
4151	ASMAtomicXchgU64(pu64, u64);
4152	#endif
4153	}
4154
4155
4156	/**
4157	* Atomically writes an unsigned 64-bit value, unordered.
4158	*
4159	* @param pu64 Pointer to the 64-bit variable to write.
4160	* @param u64 The 64-bit value to assign to *pu64.
4161	*/
4162	DECLINLINE(void) ASMAtomicUoWriteU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
4163	{
4164	Assert(!((uintptr_t)pu64 & 7));
4165	#if ARCH_BITS == 64
4166	*pu64 = u64;
4167	#else
4168	ASMAtomicXchgU64(pu64, u64);
4169	#endif
4170	}
4171
4172
4173	/**
4174	* Atomically writes a signed 64-bit value, ordered.
4175	*
4176	* @param pi64 Pointer to the 64-bit variable to write.
4177	* @param i64 The 64-bit value to assign to *pi64.
4178	*/
4179	DECLINLINE(void) ASMAtomicWriteS64(volatile int64_t RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
4180	{
4181	#if defined(RT_ARCH_ARM64)
4182	ASMAtomicWriteU64((volatile uint64_t RT_FAR *)pi64, (uint64_t)i64);
4183	#else
4184	ASMAtomicXchgS64(pi64, i64);
4185	#endif
4186	}
4187
4188
4189	/**
4190	* Atomically writes a signed 64-bit value, unordered.
4191	*
4192	* @param pi64 Pointer to the 64-bit variable to write.
4193	* @param i64 The 64-bit value to assign to *pi64.
4194	*/
4195	DECLINLINE(void) ASMAtomicUoWriteS64(volatile int64_t RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
4196	{
4197	Assert(!((uintptr_t)pi64 & 7));
4198	#if ARCH_BITS == 64
4199	*pi64 = i64;
4200	#else
4201	ASMAtomicXchgS64(pi64, i64);
4202	#endif
4203	}
4204
4205
4206	/** @def RTASM_HAVE_WRITE_U128
4207	* Defined in the target architecture supports atomic of 128-bit integers.
4208	*
4209	* The define value is zero if both ordered and unordered writes are implemented
4210	* using ASMAtomicCmpXchgU128v2(). It is 1 if unordered writes are done
4211	* natively w/o cmpxchg and 3 if both variants are done natively w/o cmpxchg.
4212	*
4213	* @note AMD64: Caller must check for cmpxchg16b support before use.
4214	* @sa RTASM_HAVE_CMP_XCHG_U128
4215	*/
4216	#if defined(RT_ARCH_ARM64) \|\| defined(DOXYGEN_RUNNING)
4217	# define RTASM_HAVE_WRITE_U128 3
4218	#elif defined(RTASM_HAVE_CMP_XCHG_U128)
4219	# define RTASM_HAVE_WRITE_U128 0
4220	#endif
4221
4222	#ifdef RTASM_HAVE_WRITE_U128
4223
4224	/**
4225	* Atomically writes an unsigned 128-bit value, ordered.
4226	*
4227	* @param pu128 Pointer to the variable to overwrite. Must be aligned
4228	* on 16 byte boundrary.
4229	* @param u64Hi The high 64 bits of the new value.
4230	* @param u64Lo The low 64 bits of the new value.
4231	*/
4232	DECLINLINE(void) ASMAtomicWriteU128v2(volatile uint128_t *pu128, const uint64_t u64Hi, const uint64_t u64Lo) RT_NOTHROW_DEF
4233	{
4234	# if !defined(__GNUC__) \|\| !defined(RT_ARCH_ARM64)
4235	RTUINT128U u128Old;
4236	# endif
4237	Assert(!((uintptr_t)pu128 & 15));
4238	# if defined(__GNUC__) && defined(RT_ARCH_ARM64)
4239	__asm__ __volatile__("Lstart_ASMAtomicWriteU128v2_%=:\n\t"
4240	# if 0 && defined(RTASM_ARM64_USE_FEAT_LSE128) /** @todo hw support? test + debug */
4241	RTASM_ARM_DMB_SY
4242	"swpp %[uValueLo], %[uValueHi], %[pMem]\n\t"
4243	# else
4244	RTASM_ARM_DMB_SY
4245	"stp %[uValueLo], %[uValueHi], %[pMem]\n\t"
4246	"dmb sy\n\t"
4247	# endif
4248	: [pMem] "+Q" (*pu128)
4249	: [uValueHi] "r" (u64Hi)
4250	, [uValueLo] "r" (u64Lo)
4251	: );
4252
4253	# else
4254	# ifdef RT_COMPILER_WITH_128BIT_INT_TYPES
4255	u128Old.u = *pu128;
4256	# else
4257	u128Old.u.Lo = pu128->Lo;
4258	u128Old.u.Hi = pu128->Hi;
4259	# endif
4260	while (!ASMAtomicCmpXchgU128v2(pu128, u64Hi, u64Lo, u128Old.s.Hi, u128Old.s.Lo, &u128Old.u))
4261	{ }
4262	# endif
4263	}
4264
4265
4266	/**
4267	* Atomically writes an unsigned 128-bit value, ordered.
4268	*
4269	* @param pu128 Pointer to the variable to overwrite. Must be aligned
4270	* on 16 byte boundrary.
4271	* @param u64Hi The high 64 bits of the new value.
4272	* @param u64Lo The low 64 bits of the new value.
4273	* @note This is ordered on AMD64.
4274	*/
4275	DECLINLINE(void) ASMAtomicUoWriteU128v2(volatile uint128_t *pu128, const uint64_t u64Hi, const uint64_t u64Lo) RT_NOTHROW_DEF
4276	{
4277	# if !defined(__GNUC__) \|\| !defined(RT_ARCH_ARM64)
4278	RTUINT128U u128Old;
4279	# endif
4280	Assert(!((uintptr_t)pu128 & 15));
4281	# if defined(__GNUC__) && defined(RT_ARCH_ARM64)
4282	__asm__ __volatile__("Lstart_ASMAtomicUoWriteU128v2_%=:\n\t"
4283	"stp %[uValueLo], %[uValueHi], %[pMem]\n\t"
4284	: [pMem] "+Q" (*pu128)
4285	: [uValueHi] "r" (u64Hi)
4286	, [uValueLo] "r" (u64Lo)
4287	: );
4288
4289	# else
4290	# ifdef RT_COMPILER_WITH_128BIT_INT_TYPES
4291	u128Old.u = *pu128;
4292	# else
4293	u128Old.u.Lo = pu128->Lo;
4294	u128Old.u.Hi = pu128->Hi;
4295	# endif
4296	while (!ASMAtomicCmpXchgU128v2(pu128, u64Hi, u64Lo, u128Old.s.Hi, u128Old.s.Lo, &u128Old.u))
4297	{ }
4298	# endif
4299	}
4300
4301
4302	/**
4303	* Atomically writes an unsigned 128-bit value, ordered.
4304	*
4305	* @param pu128 Pointer to the variable to overwrite. Must be aligned
4306	* on 16 byte boundrary.
4307	* @param u128 The the new value.
4308	*/
4309	DECLINLINE(void) ASMAtomicWriteU128(volatile uint128_t *pu128, const uint128_t u128) RT_NOTHROW_DEF
4310	{
4311	# ifdef RT_COMPILER_WITH_128BIT_INT_TYPES
4312	ASMAtomicWriteU128v2(pu128, (uint64_t)(u128 >> 64), (uint64_t)u128);
4313	# else
4314	ASMAtomicWriteU128v2(pu128, u128.Hi, u128.Lo);
4315	# endif
4316	}
4317
4318
4319	/**
4320	* Atomically writes an unsigned 128-bit value, unordered.
4321	*
4322	* @param pu128 Pointer to the variable to overwrite. Must be aligned
4323	* on 16 byte boundrary.
4324	* @param u128 The the new value.
4325	* @note This is ordered on AMD64.
4326	*/
4327	DECLINLINE(void) ASMAtomicUoWriteU128(volatile uint128_t *pu128, const uint128_t u128) RT_NOTHROW_DEF
4328	{
4329	# ifdef RT_COMPILER_WITH_128BIT_INT_TYPES
4330	ASMAtomicUoWriteU128v2(pu128, (uint64_t)(u128 >> 64), (uint64_t)u128);
4331	# else
4332	ASMAtomicUoWriteU128v2(pu128, u128.Hi, u128.Lo);
4333	# endif
4334	}
4335
4336
4337	/**
4338	* Atomically writes an unsigned 128-bit value, ordered.
4339	*
4340	* @param pu128 Pointer to the variable to overwrite. Must be aligned
4341	* on 16 byte boundrary.
4342	* @param u128 The the new value.
4343	*/
4344	DECLINLINE(void) ASMAtomicWriteU128U(volatile RTUINT128U *pu128, const RTUINT128U u128) RT_NOTHROW_DEF
4345	{
4346	ASMAtomicWriteU128v2(&pu128->u, u128.s.Hi, u128.s.Lo);
4347	}
4348
4349
4350	/**
4351	* Atomically writes an unsigned 128-bit value, unordered.
4352	*
4353	* @param pu128 Pointer to the variable to overwrite. Must be aligned
4354	* on 16 byte boundrary.
4355	* @param u128 The the new value.
4356	* @note This is ordered on AMD64.
4357	*/
4358	DECLINLINE(void) ASMAtomicUoWriteU128U(volatile RTUINT128U *pu128, const RTUINT128U u128) RT_NOTHROW_DEF
4359	{
4360	ASMAtomicUoWriteU128v2(&pu128->u, u128.s.Hi, u128.s.Lo);
4361	}
4362
4363	#endif /* RTASM_HAVE_WRITE_U128 */
4364
4365	/**
4366	* Atomically writes a size_t value, ordered.
4367	*
4368	* @param pcb Pointer to the size_t variable to write.
4369	* @param cb The value to assign to *pcb.
4370	*/
4371	DECLINLINE(void) ASMAtomicWriteZ(volatile size_t RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF
4372	{
4373	#if ARCH_BITS == 64
4374	ASMAtomicWriteU64((uint64_t volatile *)pcb, cb);
4375	#elif ARCH_BITS == 32
4376	ASMAtomicWriteU32((uint32_t volatile *)pcb, cb);
4377	#elif ARCH_BITS == 16
4378	AssertCompileSize(size_t, 2);
4379	ASMAtomicWriteU16((uint16_t volatile *)pcb, cb);
4380	#else
4381	# error "Unsupported ARCH_BITS value"
4382	#endif
4383	}
4384
4385
4386	/**
4387	* Atomically writes a size_t value, unordered.
4388	*
4389	* @param pcb Pointer to the size_t variable to write.
4390	* @param cb The value to assign to *pcb.
4391	*/
4392	DECLINLINE(void) ASMAtomicUoWriteZ(volatile size_t RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF
4393	{
4394	#if ARCH_BITS == 64
4395	ASMAtomicUoWriteU64((uint64_t volatile *)pcb, cb);
4396	#elif ARCH_BITS == 32
4397	ASMAtomicUoWriteU32((uint32_t volatile *)pcb, cb);
4398	#elif ARCH_BITS == 16
4399	AssertCompileSize(size_t, 2);
4400	ASMAtomicUoWriteU16((uint16_t volatile *)pcb, cb);
4401	#else
4402	# error "Unsupported ARCH_BITS value"
4403	#endif
4404	}
4405
4406
4407	/**
4408	* Atomically writes a boolean value, unordered.
4409	*
4410	* @param pf Pointer to the boolean variable to write.
4411	* @param f The boolean value to assign to *pf.
4412	*/
4413	DECLINLINE(void) ASMAtomicWriteBool(volatile bool RT_FAR *pf, bool f) RT_NOTHROW_DEF
4414	{
4415	ASMAtomicWriteU8((uint8_t volatile RT_FAR *)pf, f);
4416	}
4417
4418
4419	/**
4420	* Atomically writes a boolean value, unordered.
4421	*
4422	* @param pf Pointer to the boolean variable to write.
4423	* @param f The boolean value to assign to *pf.
4424	*/
4425	DECLINLINE(void) ASMAtomicUoWriteBool(volatile bool RT_FAR *pf, bool f) RT_NOTHROW_DEF
4426	{
4427	pf = f; / byte writes are atomic on x86 */
4428	}
4429
4430
4431	/**
4432	* Atomically writes a pointer value, ordered.
4433	*
4434	* @param ppv Pointer to the pointer variable to write.
4435	* @param pv The pointer value to assign to *ppv.
4436	*/
4437	DECLINLINE(void) ASMAtomicWritePtrVoid(void RT_FAR * volatile RT_FAR ppv, const void pv) RT_NOTHROW_DEF
4438	{
4439	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
4440	ASMAtomicWriteU32((volatile uint32_t RT_FAR )(void RT_FAR )ppv, (uint32_t)pv);
4441	#elif ARCH_BITS == 64
4442	ASMAtomicWriteU64((volatile uint64_t RT_FAR )(void RT_FAR )ppv, (uint64_t)pv);
4443	#else
4444	# error "ARCH_BITS is bogus"
4445	#endif
4446	}
4447
4448
4449	/**
4450	* Atomically writes a pointer value, unordered.
4451	*
4452	* @param ppv Pointer to the pointer variable to write.
4453	* @param pv The pointer value to assign to *ppv.
4454	*/
4455	DECLINLINE(void) ASMAtomicUoWritePtrVoid(void RT_FAR * volatile RT_FAR ppv, const void pv) RT_NOTHROW_DEF
4456	{
4457	#if ARCH_BITS == 32 \|\| ARCH_BITS == 16
4458	ASMAtomicUoWriteU32((volatile uint32_t RT_FAR )(void RT_FAR )ppv, (uint32_t)pv);
4459	#elif ARCH_BITS == 64
4460	ASMAtomicUoWriteU64((volatile uint64_t RT_FAR )(void RT_FAR )ppv, (uint64_t)pv);
4461	#else
4462	# error "ARCH_BITS is bogus"
4463	#endif
4464	}
4465
4466
4467	/**
4468	* Atomically writes a pointer value, ordered.
4469	*
4470	* @param ppv Pointer to the pointer variable to write.
4471	* @param pv The pointer value to assign to *ppv. If NULL use
4472	* ASMAtomicWriteNullPtr or you'll land in trouble.
4473	*
4474	* @remarks This is relatively type safe on GCC platforms when @a pv isn't
4475	* NULL.
4476	*/
4477	#ifdef __GNUC__
4478	# define ASMAtomicWritePtr(ppv, pv) \
4479	do \
4480	{ \
4481	__typeof__((ppv)) volatile RT_FAR const ppvTypeChecked = (ppv); \
4482	__typeof__(*(ppv)) const pvTypeChecked = (pv); \
4483	\
4484	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
4485	AssertCompile(sizeof(pv) == sizeof(void RT_FAR *)); \
4486	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
4487	\
4488	ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR )(ppvTypeChecked), (void RT_FAR )(pvTypeChecked)); \
4489	} while (0)
4490	#else
4491	# define ASMAtomicWritePtr(ppv, pv) \
4492	do \
4493	{ \
4494	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
4495	AssertCompile(sizeof(pv) == sizeof(void RT_FAR *)); \
4496	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
4497	\
4498	ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR )(ppv), (void RT_FAR )(pv)); \
4499	} while (0)
4500	#endif
4501
4502
4503	/**
4504	* Atomically sets a pointer to NULL, ordered.
4505	*
4506	* @param ppv Pointer to the pointer variable that should be set to NULL.
4507	*
4508	* @remarks This is relatively type safe on GCC platforms.
4509	*/
4510	#if RT_GNUC_PREREQ(4, 2)
4511	# define ASMAtomicWriteNullPtr(ppv) \
4512	do \
4513	{ \
4514	__typeof__((ppv)) const ppvTypeChecked = (ppv); \
4515	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
4516	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
4517	ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR *)(ppvTypeChecked), NULL); \
4518	} while (0)
4519	#else
4520	# define ASMAtomicWriteNullPtr(ppv) \
4521	do \
4522	{ \
4523	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
4524	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
4525	ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR *)(ppv), NULL); \
4526	} while (0)
4527	#endif
4528
4529
4530	/**
4531	* Atomically writes a pointer value, unordered.
4532	*
4533	* @returns Current *pv value
4534	* @param ppv Pointer to the pointer variable.
4535	* @param pv The pointer value to assign to *ppv. If NULL use
4536	* ASMAtomicUoWriteNullPtr or you'll land in trouble.
4537	*
4538	* @remarks This is relatively type safe on GCC platforms when @a pv isn't
4539	* NULL.
4540	*/
4541	#if RT_GNUC_PREREQ(4, 2)
4542	# define ASMAtomicUoWritePtr(ppv, pv) \
4543	do \
4544	{ \
4545	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
4546	__typeof__(*(ppv)) const pvTypeChecked = (pv); \
4547	\
4548	AssertCompile(sizeof(ppv) == sizeof(void )); \
4549	AssertCompile(sizeof(pv) == sizeof(void *)); \
4550	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
4551	\
4552	*(ppvTypeChecked) = pvTypeChecked; \
4553	} while (0)
4554	#else
4555	# define ASMAtomicUoWritePtr(ppv, pv) \
4556	do \
4557	{ \
4558	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
4559	AssertCompile(sizeof(pv) == sizeof(void RT_FAR *)); \
4560	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
4561	*(ppv) = pv; \
4562	} while (0)
4563	#endif
4564
4565
4566	/**
4567	* Atomically sets a pointer to NULL, unordered.
4568	*
4569	* @param ppv Pointer to the pointer variable that should be set to NULL.
4570	*
4571	* @remarks This is relatively type safe on GCC platforms.
4572	*/
4573	#ifdef __GNUC__
4574	# define ASMAtomicUoWriteNullPtr(ppv) \
4575	do \
4576	{ \
4577	__typeof__((ppv)) volatile const ppvTypeChecked = (ppv); \
4578	AssertCompile(sizeof(ppv) == sizeof(void )); \
4579	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
4580	*(ppvTypeChecked) = NULL; \
4581	} while (0)
4582	#else
4583	# define ASMAtomicUoWriteNullPtr(ppv) \
4584	do \
4585	{ \
4586	AssertCompile(sizeof(ppv) == sizeof(void RT_FAR )); \
4587	Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \
4588	*(ppv) = NULL; \
4589	} while (0)
4590	#endif
4591
4592
4593	/**
4594	* Atomically write a typical IPRT handle value, ordered.
4595	*
4596	* @param ph Pointer to the variable to update.
4597	* @param hNew The value to assign to *ph.
4598	*
4599	* @remarks This doesn't currently work for all handles (like RTFILE).
4600	*/
4601	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
4602	# define ASMAtomicWriteHandle(ph, hNew) \
4603	do { \
4604	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
4605	ASMAtomicWriteU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)(hNew)); \
4606	} while (0)
4607	#elif HC_ARCH_BITS == 64
4608	# define ASMAtomicWriteHandle(ph, hNew) \
4609	do { \
4610	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
4611	ASMAtomicWriteU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)(hNew)); \
4612	} while (0)
4613	#else
4614	# error HC_ARCH_BITS
4615	#endif
4616
4617
4618	/**
4619	* Atomically write a typical IPRT handle value, unordered.
4620	*
4621	* @param ph Pointer to the variable to update.
4622	* @param hNew The value to assign to *ph.
4623	*
4624	* @remarks This doesn't currently work for all handles (like RTFILE).
4625	*/
4626	#if HC_ARCH_BITS == 32 \|\| ARCH_BITS == 16
4627	# define ASMAtomicUoWriteHandle(ph, hNew) \
4628	do { \
4629	AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \
4630	ASMAtomicUoWriteU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)hNew); \
4631	} while (0)
4632	#elif HC_ARCH_BITS == 64
4633	# define ASMAtomicUoWriteHandle(ph, hNew) \
4634	do { \
4635	AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \
4636	ASMAtomicUoWriteU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)hNew); \
4637	} while (0)
4638	#else
4639	# error HC_ARCH_BITS
4640	#endif
4641
4642
4643	/**
4644	* Atomically write a value which size might differ
4645	* between platforms or compilers, ordered.
4646	*
4647	* @param pu Pointer to the variable to update.
4648	* @param uNew The value to assign to *pu.
4649	*/
4650	#define ASMAtomicWriteSize(pu, uNew) \
4651	do { \
4652	switch (sizeof(*(pu))) { \
4653	case 1: ASMAtomicWriteU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu), (uint8_t )(uNew)); break; \
4654	case 2: ASMAtomicWriteU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu), (uint16_t)(uNew)); break; \
4655	case 4: ASMAtomicWriteU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
4656	case 8: ASMAtomicWriteU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
4657	default: AssertMsgFailed(("ASMAtomicWriteSize: size %d is not supported\n", sizeof(*(pu)))); \
4658	} \
4659	} while (0)
4660
4661	/**
4662	* Atomically write a value which size might differ
4663	* between platforms or compilers, unordered.
4664	*
4665	* @param pu Pointer to the variable to update.
4666	* @param uNew The value to assign to *pu.
4667	*/
4668	#define ASMAtomicUoWriteSize(pu, uNew) \
4669	do { \
4670	switch (sizeof(*(pu))) { \
4671	case 1: ASMAtomicUoWriteU8( (volatile uint8_t RT_FAR )(void RT_FAR )(pu), (uint8_t )(uNew)); break; \
4672	case 2: ASMAtomicUoWriteU16((volatile uint16_t RT_FAR )(void RT_FAR )(pu), (uint16_t)(uNew)); break; \
4673	case 4: ASMAtomicUoWriteU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
4674	case 8: ASMAtomicUoWriteU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
4675	default: AssertMsgFailed(("ASMAtomicWriteSize: size %d is not supported\n", sizeof(*(pu)))); \
4676	} \
4677	} while (0)
4678
4679
4680
4681	/**
4682	* Atomically exchanges and adds to a 16-bit value, ordered.
4683	*
4684	* @returns The old value.
4685	* @param pu16 Pointer to the value.
4686	* @param u16 Number to add.
4687	*
4688	* @remarks Currently not implemented, just to make 16-bit code happy.
4689	* @remarks x86: Requires a 486 or later.
4690	*/
4691	RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMAtomicAddU16(uint16_t volatile RT_FAR *pu16, uint32_t u16) RT_NOTHROW_PROTO;
4692
4693
4694	/**
4695	* Atomically exchanges and adds to a 32-bit value, ordered.
4696	*
4697	* @returns The old value.
4698	* @param pu32 Pointer to the value.
4699	* @param u32 Number to add.
4700	*
4701	* @remarks x86: Requires a 486 or later.
4702	*/
4703	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4704	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicAddU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
4705	#else
4706	DECLINLINE(uint32_t) ASMAtomicAddU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
4707	{
4708	# if RT_INLINE_ASM_USES_INTRIN
4709	u32 = _InterlockedExchangeAdd((long RT_FAR *)pu32, u32);
4710	return u32;
4711
4712	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
4713	# if RT_INLINE_ASM_GNU_STYLE
4714	__asm__ __volatile__("lock; xaddl %0, %1\n\t"
4715	: "=r" (u32)
4716	, "=m" (*pu32)
4717	: "0" (u32)
4718	, "m" (*pu32)
4719	: "memory"
4720	, "cc");
4721	return u32;
4722	# else
4723	__asm
4724	{
4725	mov eax, [u32]
4726	# ifdef RT_ARCH_AMD64
4727	mov rdx, [pu32]
4728	lock xadd [rdx], eax
4729	# else
4730	mov edx, [pu32]
4731	lock xadd [edx], eax
4732	# endif
4733	mov [u32], eax
4734	}
4735	return u32;
4736	# endif
4737
4738	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4739	/* M1 benchmark: ldaddal=6907 vs dmb+ldadd=2114 vs non-lse=6249 (ps/call) */
4740	# if defined(RTASM_ARM64_USE_FEAT_LSE)
4741	uint32_t u32OldRet;
4742	__asm__ __volatile__("Lstart_ASMAtomicAddU32_%=:\n\t"
4743	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
4744	"ldaddal %w[uAddend], %w[uOldActual], %[pMem]\n\t"
4745	# else
4746	RTASM_ARM_DMB_SY
4747	"ldadd %w[uAddend], %w[uOldActual], %[pMem]\n\t"
4748	# endif
4749	: [pMem] "+Q" (*pu32)
4750	, [uOldActual] "=&r" (u32OldRet)
4751	: [uAddend] "r" (u32)
4752	: );
4753	# else
4754	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicAddU32, pu32, DMB_SY,
4755	"add %w[uNew], %w[uOld], %w[uVal]\n\t",
4756	"add %[uNew], %[uOld], %[uVal]\n\t",
4757	[uVal] "r" (u32));
4758	# endif
4759	return u32OldRet;
4760
4761	# else
4762	# error "Port me"
4763	# endif
4764	}
4765	#endif
4766
4767
4768	/**
4769	* Atomically exchanges and adds to a signed 32-bit value, ordered.
4770	*
4771	* @returns The old value.
4772	* @param pi32 Pointer to the value.
4773	* @param i32 Number to add.
4774	*
4775	* @remarks x86: Requires a 486 or later.
4776	*/
4777	DECLINLINE(int32_t) ASMAtomicAddS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
4778	{
4779	return (int32_t)ASMAtomicAddU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
4780	}
4781
4782
4783	/**
4784	* Atomically exchanges and adds to a 64-bit value, ordered.
4785	*
4786	* @returns The old value.
4787	* @param pu64 Pointer to the value.
4788	* @param u64 Number to add.
4789	*
4790	* @remarks x86: Requires a Pentium or later.
4791	*/
4792	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
4793	DECLASM(uint64_t) ASMAtomicAddU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
4794	#else
4795	DECLINLINE(uint64_t) ASMAtomicAddU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
4796	{
4797	# if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64)
4798	u64 = _InterlockedExchangeAdd64((__int64 RT_FAR *)pu64, u64);
4799	return u64;
4800
4801	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
4802	__asm__ __volatile__("lock; xaddq %0, %1\n\t"
4803	: "=r" (u64)
4804	, "=m" (*pu64)
4805	: "0" (u64)
4806	, "m" (*pu64)
4807	: "memory"
4808	, "cc");
4809	return u64;
4810
4811	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
4812	# if defined(RTASM_ARM64_USE_FEAT_LSE)
4813	uint64_t u64OldRet;
4814	__asm__ __volatile__("Lstart_ASMAtomicAddU64_%=:\n\t"
4815	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
4816	"ldaddal %[uAddend], %[uOldActual], %[pMem]\n\t"
4817	# else
4818	RTASM_ARM_DMB_SY
4819	"ldadd %[uAddend], %[uOldActual], %[pMem]\n\t"
4820	# endif
4821	: [pMem] "+Q" (*pu64)
4822	, [uOldActual] "=&r" (u64OldRet)
4823	: [uAddend] "r" (u64)
4824	: );
4825	# else
4826	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_64(ASMAtomicAddU64, pu64, DMB_SY,
4827	"add %[uNew], %[uOld], %[uVal]\n\t"
4828	,
4829	"add %[uNew], %[uOld], %[uVal]\n\t"
4830	"adc %H[uNew], %H[uOld], %H[uVal]\n\t",
4831	[uVal] "r" (u64));
4832	# endif
4833	return u64OldRet;
4834
4835	# else
4836	uint64_t u64Old;
4837	for (;;)
4838	{
4839	uint64_t u64New;
4840	u64Old = ASMAtomicUoReadU64(pu64);
4841	u64New = u64Old + u64;
4842	if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old))
4843	break;
4844	ASMNopPause();
4845	}
4846	return u64Old;
4847	# endif
4848	}
4849	#endif
4850
4851
4852	/**
4853	* Atomically exchanges and adds to a signed 64-bit value, ordered.
4854	*
4855	* @returns The old value.
4856	* @param pi64 Pointer to the value.
4857	* @param i64 Number to add.
4858	*
4859	* @remarks x86: Requires a Pentium or later.
4860	*/
4861	DECLINLINE(int64_t) ASMAtomicAddS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
4862	{
4863	return (int64_t)ASMAtomicAddU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64);
4864	}
4865
4866
4867	/**
4868	* Atomically exchanges and adds to a size_t value, ordered.
4869	*
4870	* @returns The old value.
4871	* @param pcb Pointer to the size_t value.
4872	* @param cb Number to add.
4873	*/
4874	DECLINLINE(size_t) ASMAtomicAddZ(size_t volatile RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF
4875	{
4876	#if ARCH_BITS == 64
4877	AssertCompileSize(size_t, 8);
4878	return ASMAtomicAddU64((uint64_t volatile RT_FAR *)pcb, cb);
4879	#elif ARCH_BITS == 32
4880	AssertCompileSize(size_t, 4);
4881	return ASMAtomicAddU32((uint32_t volatile RT_FAR *)pcb, cb);
4882	#elif ARCH_BITS == 16
4883	AssertCompileSize(size_t, 2);
4884	return ASMAtomicAddU16((uint16_t volatile RT_FAR *)pcb, cb);
4885	#else
4886	# error "Unsupported ARCH_BITS value"
4887	#endif
4888	}
4889
4890
4891	/**
4892	* Atomically exchanges and adds a value which size might differ between
4893	* platforms or compilers, ordered.
4894	*
4895	* @param pu Pointer to the variable to update.
4896	* @param uNew The value to add to *pu.
4897	* @param puOld Where to store the old value.
4898	*/
4899	#define ASMAtomicAddSize(pu, uNew, puOld) \
4900	do { \
4901	switch (sizeof(*(pu))) { \
4902	case 4: (uint32_t )(puOld) = ASMAtomicAddU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
4903	case 8: (uint64_t )(puOld) = ASMAtomicAddU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
4904	default: AssertMsgFailed(("ASMAtomicAddSize: size %d is not supported\n", sizeof(*(pu)))); \
4905	} \
4906	} while (0)
4907
4908
4909
4910	/**
4911	* Atomically exchanges and subtracts to an unsigned 16-bit value, ordered.
4912	*
4913	* @returns The old value.
4914	* @param pu16 Pointer to the value.
4915	* @param u16 Number to subtract.
4916	*
4917	* @remarks x86: Requires a 486 or later.
4918	*/
4919	DECLINLINE(uint16_t) ASMAtomicSubU16(uint16_t volatile RT_FAR *pu16, uint32_t u16) RT_NOTHROW_DEF
4920	{
4921	return ASMAtomicAddU16(pu16, (uint16_t)-(int16_t)u16);
4922	}
4923
4924
4925	/**
4926	* Atomically exchanges and subtracts to a signed 16-bit value, ordered.
4927	*
4928	* @returns The old value.
4929	* @param pi16 Pointer to the value.
4930	* @param i16 Number to subtract.
4931	*
4932	* @remarks x86: Requires a 486 or later.
4933	*/
4934	DECLINLINE(int16_t) ASMAtomicSubS16(int16_t volatile RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF
4935	{
4936	return (int16_t)ASMAtomicAddU16((uint16_t volatile RT_FAR *)pi16, (uint16_t)-i16);
4937	}
4938
4939
4940	/**
4941	* Atomically exchanges and subtracts to an unsigned 32-bit value, ordered.
4942	*
4943	* @returns The old value.
4944	* @param pu32 Pointer to the value.
4945	* @param u32 Number to subtract.
4946	*
4947	* @remarks x86: Requires a 486 or later.
4948	*/
4949	DECLINLINE(uint32_t) ASMAtomicSubU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
4950	{
4951	return ASMAtomicAddU32(pu32, (uint32_t)-(int32_t)u32);
4952	}
4953
4954
4955	/**
4956	* Atomically exchanges and subtracts to a signed 32-bit value, ordered.
4957	*
4958	* @returns The old value.
4959	* @param pi32 Pointer to the value.
4960	* @param i32 Number to subtract.
4961	*
4962	* @remarks x86: Requires a 486 or later.
4963	*/
4964	DECLINLINE(int32_t) ASMAtomicSubS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
4965	{
4966	return (int32_t)ASMAtomicAddU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)-i32);
4967	}
4968
4969
4970	/**
4971	* Atomically exchanges and subtracts to an unsigned 64-bit value, ordered.
4972	*
4973	* @returns The old value.
4974	* @param pu64 Pointer to the value.
4975	* @param u64 Number to subtract.
4976	*
4977	* @remarks x86: Requires a Pentium or later.
4978	*/
4979	DECLINLINE(uint64_t) ASMAtomicSubU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
4980	{
4981	return ASMAtomicAddU64(pu64, (uint64_t)-(int64_t)u64);
4982	}
4983
4984
4985	/**
4986	* Atomically exchanges and subtracts to a signed 64-bit value, ordered.
4987	*
4988	* @returns The old value.
4989	* @param pi64 Pointer to the value.
4990	* @param i64 Number to subtract.
4991	*
4992	* @remarks x86: Requires a Pentium or later.
4993	*/
4994	DECLINLINE(int64_t) ASMAtomicSubS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
4995	{
4996	return (int64_t)ASMAtomicAddU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)-i64);
4997	}
4998
4999
5000	/**
5001	* Atomically exchanges and subtracts to a size_t value, ordered.
5002	*
5003	* @returns The old value.
5004	* @param pcb Pointer to the size_t value.
5005	* @param cb Number to subtract.
5006	*
5007	* @remarks x86: Requires a 486 or later.
5008	*/
5009	DECLINLINE(size_t) ASMAtomicSubZ(size_t volatile RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF
5010	{
5011	#if ARCH_BITS == 64
5012	return ASMAtomicSubU64((uint64_t volatile RT_FAR *)pcb, cb);
5013	#elif ARCH_BITS == 32
5014	return ASMAtomicSubU32((uint32_t volatile RT_FAR *)pcb, cb);
5015	#elif ARCH_BITS == 16
5016	AssertCompileSize(size_t, 2);
5017	return ASMAtomicSubU16((uint16_t volatile RT_FAR *)pcb, cb);
5018	#else
5019	# error "Unsupported ARCH_BITS value"
5020	#endif
5021	}
5022
5023
5024	/**
5025	* Atomically exchanges and subtracts a value which size might differ between
5026	* platforms or compilers, ordered.
5027	*
5028	* @param pu Pointer to the variable to update.
5029	* @param uNew The value to subtract to *pu.
5030	* @param puOld Where to store the old value.
5031	*
5032	* @remarks x86: Requires a 486 or later.
5033	*/
5034	#define ASMAtomicSubSize(pu, uNew, puOld) \
5035	do { \
5036	switch (sizeof(*(pu))) { \
5037	case 4: (uint32_t RT_FAR )(puOld) = ASMAtomicSubU32((volatile uint32_t RT_FAR )(void RT_FAR )(pu), (uint32_t)(uNew)); break; \
5038	case 8: (uint64_t RT_FAR )(puOld) = ASMAtomicSubU64((volatile uint64_t RT_FAR )(void RT_FAR )(pu), (uint64_t)(uNew)); break; \
5039	default: AssertMsgFailed(("ASMAtomicSubSize: size %d is not supported\n", sizeof(*(pu)))); \
5040	} \
5041	} while (0)
5042
5043
5044
5045	/**
5046	* Atomically increment a 16-bit value, ordered.
5047	*
5048	* @returns The new value.
5049	* @param pu16 Pointer to the value to increment.
5050	* @remarks Not implemented. Just to make 16-bit code happy.
5051	*
5052	* @remarks x86: Requires a 486 or later.
5053	*/
5054	RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMAtomicIncU16(uint16_t volatile RT_FAR *pu16) RT_NOTHROW_PROTO;
5055
5056
5057	/**
5058	* Atomically increment a 32-bit value, ordered.
5059	*
5060	* @returns The new value.
5061	* @param pu32 Pointer to the value to increment.
5062	*
5063	* @remarks x86: Requires a 486 or later.
5064	*/
5065	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
5066	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO;
5067	#else
5068	DECLINLINE(uint32_t) ASMAtomicIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF
5069	{
5070	# if RT_INLINE_ASM_USES_INTRIN
5071	return (uint32_t)_InterlockedIncrement((long RT_FAR *)pu32);
5072
5073	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5074	# if RT_INLINE_ASM_GNU_STYLE
5075	uint32_t u32;
5076	__asm__ __volatile__("lock; xaddl %0, %1\n\t"
5077	: "=r" (u32)
5078	, "=m" (*pu32)
5079	: "0" (1)
5080	, "m" (*pu32)
5081	: "memory"
5082	, "cc");
5083	return u32+1;
5084	# else
5085	__asm
5086	{
5087	mov eax, 1
5088	# ifdef RT_ARCH_AMD64
5089	mov rdx, [pu32]
5090	lock xadd [rdx], eax
5091	# else
5092	mov edx, [pu32]
5093	lock xadd [edx], eax
5094	# endif
5095	mov u32, eax
5096	}
5097	return u32+1;
5098	# endif
5099
5100	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5101	/* M1 benchmark: ldaddal=6887 vs dmb+ldadd=2117 vs non-lse=6247 (ps/call) */
5102	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5103	uint32_t u32NewRet;
5104	__asm__ __volatile__("Lstart_ASMAtomicIncU32_%=:\n\t"
5105	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5106	"ldaddal %w[uAddend], %w[uNewRet], %[pMem]\n\t"
5107	# else
5108	RTASM_ARM_DMB_SY
5109	"ldadd %w[uAddend], %w[uNewRet], %[pMem]\n\t"
5110	# endif
5111	"add %w[uNewRet], %w[uNewRet], #1\n\t"
5112	: [pMem] "+Q" (*pu32)
5113	, [uNewRet] "=&r" (u32NewRet)
5114	: [uAddend] "r" ((uint32_t)1)
5115	: );
5116	# else
5117	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicIncU32, pu32, DMB_SY,
5118	"add %w[uNew], %w[uNew], #1\n\t",
5119	"add %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */,
5120	"X" (0) /* dummy */);
5121	# endif
5122	return u32NewRet;
5123
5124	# else
5125	return ASMAtomicAddU32(pu32, 1) + 1;
5126	# endif
5127	}
5128	#endif
5129
5130
5131	/**
5132	* Atomically increment a signed 32-bit value, ordered.
5133	*
5134	* @returns The new value.
5135	* @param pi32 Pointer to the value to increment.
5136	*
5137	* @remarks x86: Requires a 486 or later.
5138	*/
5139	DECLINLINE(int32_t) ASMAtomicIncS32(int32_t volatile RT_FAR *pi32) RT_NOTHROW_DEF
5140	{
5141	return (int32_t)ASMAtomicIncU32((uint32_t volatile RT_FAR *)pi32);
5142	}
5143
5144
5145	/**
5146	* Atomically increment a 64-bit value, ordered.
5147	*
5148	* @returns The new value.
5149	* @param pu64 Pointer to the value to increment.
5150	*
5151	* @remarks x86: Requires a Pentium or later.
5152	*/
5153	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
5154	DECLASM(uint64_t) ASMAtomicIncU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_PROTO;
5155	#else
5156	DECLINLINE(uint64_t) ASMAtomicIncU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_DEF
5157	{
5158	# if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64)
5159	return (uint64_t)_InterlockedIncrement64((__int64 RT_FAR *)pu64);
5160
5161	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
5162	uint64_t u64;
5163	__asm__ __volatile__("lock; xaddq %0, %1\n\t"
5164	: "=r" (u64)
5165	, "=m" (*pu64)
5166	: "0" (1)
5167	, "m" (*pu64)
5168	: "memory"
5169	, "cc");
5170	return u64 + 1;
5171
5172	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5173	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5174	uint64_t u64NewRet;
5175	__asm__ __volatile__("Lstart_ASMAtomicIncU64_%=:\n\t"
5176	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5177	"ldaddal %[uAddend], %[uNewRet], %[pMem]\n\t"
5178	# else
5179	RTASM_ARM_DMB_SY
5180	"ldadd %[uAddend], %[uNewRet], %[pMem]\n\t"
5181	# endif
5182	"add %[uNewRet], %[uNewRet], #1\n\t"
5183	: [pMem] "+Q" (*pu64)
5184	, [uNewRet] "=&r" (u64NewRet)
5185	: [uAddend] "r" ((uint64_t)1)
5186	: );
5187	# else
5188	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicIncU64, pu64, DMB_SY,
5189	"add %[uNew], %[uNew], #1\n\t"
5190	,
5191	"add %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */
5192	"adc %H[uNew], %H[uNew], %[uZeroVal]\n\t",
5193	RTASM_ARM_PICK_6432("X" (0) /* dummy */, [uZeroVal] "r" (0)) );
5194	# endif
5195	return u64NewRet;
5196
5197	# else
5198	return ASMAtomicAddU64(pu64, 1) + 1;
5199	# endif
5200	}
5201	#endif
5202
5203
5204	/**
5205	* Atomically increment a signed 64-bit value, ordered.
5206	*
5207	* @returns The new value.
5208	* @param pi64 Pointer to the value to increment.
5209	*
5210	* @remarks x86: Requires a Pentium or later.
5211	*/
5212	DECLINLINE(int64_t) ASMAtomicIncS64(int64_t volatile RT_FAR *pi64) RT_NOTHROW_DEF
5213	{
5214	return (int64_t)ASMAtomicIncU64((uint64_t volatile RT_FAR *)pi64);
5215	}
5216
5217
5218	/**
5219	* Atomically increment a size_t value, ordered.
5220	*
5221	* @returns The new value.
5222	* @param pcb Pointer to the value to increment.
5223	*
5224	* @remarks x86: Requires a 486 or later.
5225	*/
5226	DECLINLINE(size_t) ASMAtomicIncZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF
5227	{
5228	#if ARCH_BITS == 64
5229	return ASMAtomicIncU64((uint64_t volatile RT_FAR *)pcb);
5230	#elif ARCH_BITS == 32
5231	return ASMAtomicIncU32((uint32_t volatile RT_FAR *)pcb);
5232	#elif ARCH_BITS == 16
5233	return ASMAtomicIncU16((uint16_t volatile RT_FAR *)pcb);
5234	#else
5235	# error "Unsupported ARCH_BITS value"
5236	#endif
5237	}
5238
5239
5240
5241	/**
5242	* Atomically decrement an unsigned 32-bit value, ordered.
5243	*
5244	* @returns The new value.
5245	* @param pu16 Pointer to the value to decrement.
5246	* @remarks Not implemented. Just to make 16-bit code happy.
5247	*
5248	* @remarks x86: Requires a 486 or later.
5249	*/
5250	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicDecU16(uint16_t volatile RT_FAR *pu16) RT_NOTHROW_PROTO;
5251
5252
5253	/**
5254	* Atomically decrement an unsigned 32-bit value, ordered.
5255	*
5256	* @returns The new value.
5257	* @param pu32 Pointer to the value to decrement.
5258	*
5259	* @remarks x86: Requires a 486 or later.
5260	*/
5261	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
5262	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO;
5263	#else
5264	DECLINLINE(uint32_t) ASMAtomicDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF
5265	{
5266	# if RT_INLINE_ASM_USES_INTRIN
5267	return (uint32_t)_InterlockedDecrement((long RT_FAR *)pu32);
5268
5269	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
5270	# if RT_INLINE_ASM_GNU_STYLE
5271	uint32_t u32;
5272	__asm__ __volatile__("lock; xaddl %0, %1\n\t"
5273	: "=r" (u32)
5274	, "=m" (*pu32)
5275	: "0" (-1)
5276	, "m" (*pu32)
5277	: "memory"
5278	, "cc");
5279	return u32-1;
5280	# else
5281	uint32_t u32;
5282	__asm
5283	{
5284	mov eax, -1
5285	# ifdef RT_ARCH_AMD64
5286	mov rdx, [pu32]
5287	lock xadd [rdx], eax
5288	# else
5289	mov edx, [pu32]
5290	lock xadd [edx], eax
5291	# endif
5292	mov u32, eax
5293	}
5294	return u32-1;
5295	# endif
5296
5297	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5298	/* M1 benchmark: ldaddal=6887 vs dmb+ldadd=2120 vs non-lse=6260 (ps/call) */
5299	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5300	uint32_t u32NewRet;
5301	__asm__ __volatile__("Lstart_ASMAtomicDecU32_%=:\n\t"
5302	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5303	"ldaddal %w[uAddend], %w[uNewRet], %[pMem]\n\t"
5304	# else
5305	RTASM_ARM_DMB_SY
5306	"ldadd %w[uAddend], %w[uNewRet], %[pMem]\n\t"
5307	# endif
5308	"sub %w[uNewRet], %w[uNewRet], #1\n\t"
5309	: [pMem] "+Q" (*pu32)
5310	, [uNewRet] "=&r" (u32NewRet)
5311	: [uAddend] "r" (~(uint32_t)0)
5312	: );
5313	# else
5314	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicDecU32, pu32, DMB_SY,
5315	"sub %w[uNew], %w[uNew], #1\n\t",
5316	"sub %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */,
5317	"X" (0) /* dummy */);
5318	# endif
5319	return u32NewRet;
5320
5321	# else
5322	return ASMAtomicSubU32(pu32, 1) - (uint32_t)1;
5323	# endif
5324	}
5325	#endif
5326
5327
5328	/**
5329	* Atomically decrement a signed 32-bit value, ordered.
5330	*
5331	* @returns The new value.
5332	* @param pi32 Pointer to the value to decrement.
5333	*
5334	* @remarks x86: Requires a 486 or later.
5335	*/
5336	DECLINLINE(int32_t) ASMAtomicDecS32(int32_t volatile RT_FAR *pi32) RT_NOTHROW_DEF
5337	{
5338	return (int32_t)ASMAtomicDecU32((uint32_t volatile RT_FAR *)pi32);
5339	}
5340
5341
5342	/**
5343	* Atomically decrement an unsigned 64-bit value, ordered.
5344	*
5345	* @returns The new value.
5346	* @param pu64 Pointer to the value to decrement.
5347	*
5348	* @remarks x86: Requires a Pentium or later.
5349	*/
5350	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
5351	RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicDecU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_PROTO;
5352	#else
5353	DECLINLINE(uint64_t) ASMAtomicDecU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_DEF
5354	{
5355	# if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64)
5356	return (uint64_t)_InterlockedDecrement64((__int64 volatile RT_FAR *)pu64);
5357
5358	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
5359	uint64_t u64;
5360	__asm__ __volatile__("lock; xaddq %q0, %1\n\t"
5361	: "=r" (u64)
5362	, "=m" (*pu64)
5363	: "0" (~(uint64_t)0)
5364	, "m" (*pu64)
5365	: "memory"
5366	, "cc");
5367	return u64-1;
5368
5369	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5370	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5371	uint64_t u64NewRet;
5372	__asm__ __volatile__("Lstart_ASMAtomicDecU64_%=:\n\t"
5373	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5374	"ldaddal %[uAddend], %[uNewRet], %[pMem]\n\t"
5375	# else
5376	RTASM_ARM_DMB_SY
5377	"ldadd %[uAddend], %[uNewRet], %[pMem]\n\t"
5378	# endif
5379	"sub %[uNewRet], %[uNewRet], #1\n\t"
5380	: [pMem] "+Q" (*pu64)
5381	, [uNewRet] "=&r" (u64NewRet)
5382	: [uAddend] "r" (~(uint64_t)0)
5383	: );
5384	# else
5385	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicDecU64, pu64, DMB_SY,
5386	"sub %[uNew], %[uNew], #1\n\t"
5387	,
5388	"sub %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */
5389	"sbc %H[uNew], %H[uNew], %[uZeroVal]\n\t",
5390	RTASM_ARM_PICK_6432("X" (0) /* dummy */, [uZeroVal] "r" (0)) );
5391	# endif
5392	return u64NewRet;
5393
5394	# else
5395	return ASMAtomicAddU64(pu64, UINT64_MAX) - 1;
5396	# endif
5397	}
5398	#endif
5399
5400
5401	/**
5402	* Atomically decrement a signed 64-bit value, ordered.
5403	*
5404	* @returns The new value.
5405	* @param pi64 Pointer to the value to decrement.
5406	*
5407	* @remarks x86: Requires a Pentium or later.
5408	*/
5409	DECLINLINE(int64_t) ASMAtomicDecS64(int64_t volatile RT_FAR *pi64) RT_NOTHROW_DEF
5410	{
5411	return (int64_t)ASMAtomicDecU64((uint64_t volatile RT_FAR *)pi64);
5412	}
5413
5414
5415	/**
5416	* Atomically decrement a size_t value, ordered.
5417	*
5418	* @returns The new value.
5419	* @param pcb Pointer to the value to decrement.
5420	*
5421	* @remarks x86: Requires a 486 or later.
5422	*/
5423	DECLINLINE(size_t) ASMAtomicDecZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF
5424	{
5425	#if ARCH_BITS == 64
5426	return ASMAtomicDecU64((uint64_t volatile RT_FAR *)pcb);
5427	#elif ARCH_BITS == 32
5428	return ASMAtomicDecU32((uint32_t volatile RT_FAR *)pcb);
5429	#elif ARCH_BITS == 16
5430	return ASMAtomicDecU16((uint16_t volatile RT_FAR *)pcb);
5431	#else
5432	# error "Unsupported ARCH_BITS value"
5433	#endif
5434	}
5435
5436
5437	/**
5438	* Atomically Or an unsigned 32-bit value, ordered.
5439	*
5440	* @param pu32 Pointer to the pointer variable to OR u32 with.
5441	* @param u32 The value to OR *pu32 with.
5442	*
5443	* @remarks x86: Requires a 386 or later.
5444	*/
5445	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
5446	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
5447	#else
5448	DECLINLINE(void) ASMAtomicOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5449	{
5450	# if RT_INLINE_ASM_USES_INTRIN
5451	_InterlockedOr((long volatile RT_FAR *)pu32, (long)u32);
5452
5453	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5454	# if RT_INLINE_ASM_GNU_STYLE
5455	__asm__ __volatile__("lock; orl %1, %0\n\t"
5456	: "=m" (*pu32)
5457	: "ir" (u32)
5458	, "m" (*pu32)
5459	: "cc");
5460	# else
5461	__asm
5462	{
5463	mov eax, [u32]
5464	# ifdef RT_ARCH_AMD64
5465	mov rdx, [pu32]
5466	lock or [rdx], eax
5467	# else
5468	mov edx, [pu32]
5469	lock or [edx], eax
5470	# endif
5471	}
5472	# endif
5473
5474	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5475	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5476	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5477	uint32_t u32Spill;
5478	__asm__ __volatile__("Lstart_ASMAtomicOrU32_%=:\n\t"
5479	"ldsetal %w[fBitsToSet], %w[uSpill], %[pMem]\n\t"
5480	: [pMem] "+Q" (*pu32)
5481	, [uSpill] "=&r" (u32Spill)
5482	: [fBitsToSet] "r" (u32)
5483	: );
5484	# else
5485	__asm__ __volatile__("Lstart_ASMAtomicOrU32_%=:\n\t"
5486	RTASM_ARM_DMB_SY
5487	"stset %w[fBitsToSet], %[pMem]\n\t"
5488	: [pMem] "+Q" (*pu32)
5489	: [fBitsToSet] "r" (u32)
5490	: );
5491	# endif
5492	# else
5493	/* For more on Orr see https://en.wikipedia.org/wiki/Orr_(Catch-22) ;-) */
5494	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicOr32, pu32, DMB_SY,
5495	"orr %w[uNew], %w[uNew], %w[uVal]\n\t",
5496	"orr %[uNew], %[uNew], %[uVal]\n\t",
5497	[uVal] "r" (u32));
5498
5499	# endif
5500	# else
5501	# error "Port me"
5502	# endif
5503	}
5504	#endif
5505
5506
5507	/**
5508	* Atomically OR an unsigned 32-bit value, ordered, extended version (for bitmap
5509	* fallback).
5510	*
5511	* @returns Old value.
5512	* @param pu32 Pointer to the variable to OR @a u32 with.
5513	* @param u32 The value to OR @a *pu32 with.
5514	*/
5515	DECLINLINE(uint32_t) ASMAtomicOrExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5516	{
5517	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5518	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5519	uint32_t u32OldRet;
5520	__asm__ __volatile__("Lstart_ASMAtomicOrExU32_%=:\n\t"
5521	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5522	"ldsetal %w[fBitsToSet], %w[uOldRet], %[pMem]\n\t"
5523	# else
5524	RTASM_ARM_DMB_SY
5525	"ldset %w[fBitsToSet], %w[uOldRet], %[pMem]\n\t"
5526	# endif
5527	: [pMem] "+Q" (*pu32)
5528	, [uOldRet] "=&r" (u32OldRet)
5529	: [fBitsToSet] "r" (u32)
5530	: );
5531	# else
5532	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicOrEx32, pu32, DMB_SY,
5533	"orr %w[uNew], %w[uOld], %w[uVal]\n\t",
5534	"orr %[uNew], %[uOld], %[uVal]\n\t",
5535	[uVal] "r" (u32));
5536	# endif
5537	return u32OldRet;
5538
5539	#else
5540	uint32_t u32RetOld = ASMAtomicUoReadU32(pu32);
5541	uint32_t u32New;
5542	do
5543	u32New = u32RetOld \| u32;
5544	while (!ASMAtomicCmpXchgExU32(pu32, u32New, u32RetOld, &u32RetOld));
5545	return u32RetOld;
5546	#endif
5547	}
5548
5549
5550	/**
5551	* Atomically Or a signed 32-bit value, ordered.
5552	*
5553	* @param pi32 Pointer to the pointer variable to OR u32 with.
5554	* @param i32 The value to OR *pu32 with.
5555	*
5556	* @remarks x86: Requires a 386 or later.
5557	*/
5558	DECLINLINE(void) ASMAtomicOrS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
5559	{
5560	ASMAtomicOrU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
5561	}
5562
5563
5564	/**
5565	* Atomically Or an unsigned 64-bit value, ordered.
5566	*
5567	* @param pu64 Pointer to the pointer variable to OR u64 with.
5568	* @param u64 The value to OR *pu64 with.
5569	*
5570	* @remarks x86: Requires a Pentium or later.
5571	*/
5572	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
5573	DECLASM(void) ASMAtomicOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
5574	#else
5575	DECLINLINE(void) ASMAtomicOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
5576	{
5577	# if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64)
5578	_InterlockedOr64((__int64 volatile RT_FAR *)pu64, (__int64)u64);
5579
5580	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
5581	__asm__ __volatile__("lock; orq %1, %q0\n\t"
5582	: "=m" (*pu64)
5583	: "r" (u64)
5584	, "m" (*pu64)
5585	: "cc");
5586
5587	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5588	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5589	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5590	uint64_t u64Spill;
5591	__asm__ __volatile__("Lstart_ASMAtomicOrU64_%=:\n\t"
5592	"ldsetal %[fBitsToSet], %[uSpill], %[pMem]\n\t"
5593	: [pMem] "+Q" (*pu64)
5594	, [uSpill] "=&r" (u64Spill)
5595	: [fBitsToSet] "r" (u64)
5596	: );
5597	# else
5598	__asm__ __volatile__("Lstart_ASMAtomicOrU64_%=:\n\t"
5599	RTASM_ARM_DMB_SY
5600	"stset %[fBitsToSet], %[pMem]\n\t"
5601	: [pMem] "+Q" (*pu64)
5602	: [fBitsToSet] "r" (u64)
5603	: );
5604	# endif
5605	# else
5606	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicOrU64, pu64, DMB_SY,
5607	"orr %[uNew], %[uNew], %[uVal]\n\t"
5608	,
5609	"orr %[uNew], %[uNew], %[uVal]\n\t"
5610	"orr %H[uNew], %H[uNew], %H[uVal]\n\t",
5611	[uVal] "r" (u64));
5612	# endif
5613
5614	# else
5615	for (;;)
5616	{
5617	uint64_t u64Old = ASMAtomicUoReadU64(pu64);
5618	uint64_t u64New = u64Old \| u64;
5619	if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old))
5620	break;
5621	ASMNopPause();
5622	}
5623	# endif
5624	}
5625	#endif
5626
5627
5628	/**
5629	* Atomically Or a signed 64-bit value, ordered.
5630	*
5631	* @param pi64 Pointer to the pointer variable to OR u64 with.
5632	* @param i64 The value to OR *pu64 with.
5633	*
5634	* @remarks x86: Requires a Pentium or later.
5635	*/
5636	DECLINLINE(void) ASMAtomicOrS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
5637	{
5638	ASMAtomicOrU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64);
5639	}
5640
5641
5642	/**
5643	* Atomically And an unsigned 32-bit value, ordered.
5644	*
5645	* @param pu32 Pointer to the pointer variable to AND u32 with.
5646	* @param u32 The value to AND *pu32 with.
5647	*
5648	* @remarks x86: Requires a 386 or later.
5649	*/
5650	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
5651	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
5652	#else
5653	DECLINLINE(void) ASMAtomicAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5654	{
5655	# if RT_INLINE_ASM_USES_INTRIN
5656	_InterlockedAnd((long volatile RT_FAR *)pu32, u32);
5657
5658	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5659	# if RT_INLINE_ASM_GNU_STYLE
5660	__asm__ __volatile__("lock; andl %1, %0\n\t"
5661	: "=m" (*pu32)
5662	: "ir" (u32)
5663	, "m" (*pu32)
5664	: "cc");
5665	# else
5666	__asm
5667	{
5668	mov eax, [u32]
5669	# ifdef RT_ARCH_AMD64
5670	mov rdx, [pu32]
5671	lock and [rdx], eax
5672	# else
5673	mov edx, [pu32]
5674	lock and [edx], eax
5675	# endif
5676	}
5677	# endif
5678
5679	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5680	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5681	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5682	uint32_t u32Spill;
5683	__asm__ __volatile__("Lstart_ASMAtomicAndU32_%=:\n\t"
5684	"ldclral %w[fBitsToClear], %w[uSpill], %[pMem]\n\t"
5685	: [pMem] "+Q" (*pu32)
5686	, [uSpill] "=&r" (u32Spill)
5687	: [fBitsToClear] "r" (~u32)
5688	: );
5689	# else
5690	__asm__ __volatile__("Lstart_ASMAtomicAndU32_%=:\n\t"
5691	RTASM_ARM_DMB_SY
5692	"stclr %w[fBitsToClear], %[pMem]\n\t"
5693	: [pMem] "+Q" (*pu32)
5694	: [fBitsToClear] "r" (~u32)
5695	: );
5696	# endif
5697	# else
5698	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicAnd32, pu32, DMB_SY,
5699	"and %w[uNew], %w[uNew], %w[uVal]\n\t",
5700	"and %[uNew], %[uNew], %[uVal]\n\t",
5701	[uVal] "r" (u32));
5702
5703	# endif
5704	# else
5705	# error "Port me"
5706	# endif
5707	}
5708	#endif
5709
5710
5711	/**
5712	* Atomically AND an unsigned 32-bit value, ordered, extended version.
5713	*
5714	* @returns Old value.
5715	* @param pu32 Pointer to the variable to AND @a u32 with.
5716	* @param u32 The value to AND @a *pu32 with.
5717	*/
5718	DECLINLINE(uint32_t) ASMAtomicAndExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5719	{
5720	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5721	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5722	uint32_t u32OldRet;
5723	__asm__ __volatile__("Lstart_ASMAtomicAndExU32_%=:\n\t"
5724	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5725	"ldclral %w[fBitsToClear], %w[uOldRet], %[pMem]\n\t"
5726	# else
5727	RTASM_ARM_DMB_SY
5728	"ldclr %w[fBitsToClear], %w[uOldRet], %[pMem]\n\t"
5729	# endif
5730	: [pMem] "+Q" (*pu32)
5731	, [uOldRet] "=&r" (u32OldRet)
5732	: [fBitsToClear] "r" (~u32)
5733	: );
5734	# else
5735	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicAndEx32, pu32, DMB_SY,
5736	"and %w[uNew], %w[uOld], %w[uVal]\n\t",
5737	"and %[uNew], %[uOld], %[uVal]\n\t",
5738	[uVal] "r" (u32));
5739	# endif
5740	return u32OldRet;
5741
5742	#else
5743	uint32_t u32RetOld = ASMAtomicUoReadU32(pu32);
5744	uint32_t u32New;
5745	do
5746	u32New = u32RetOld & u32;
5747	while (!ASMAtomicCmpXchgExU32(pu32, u32New, u32RetOld, &u32RetOld));
5748	return u32RetOld;
5749	#endif
5750	}
5751
5752
5753	/**
5754	* Atomically And a signed 32-bit value, ordered.
5755	*
5756	* @param pi32 Pointer to the pointer variable to AND i32 with.
5757	* @param i32 The value to AND *pi32 with.
5758	*
5759	* @remarks x86: Requires a 386 or later.
5760	*/
5761	DECLINLINE(void) ASMAtomicAndS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
5762	{
5763	ASMAtomicAndU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
5764	}
5765
5766
5767	/**
5768	* Atomically And an unsigned 64-bit value, ordered.
5769	*
5770	* @param pu64 Pointer to the pointer variable to AND u64 with.
5771	* @param u64 The value to AND *pu64 with.
5772	*
5773	* @remarks x86: Requires a Pentium or later.
5774	*/
5775	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
5776	DECLASM(void) ASMAtomicAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
5777	#else
5778	DECLINLINE(void) ASMAtomicAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
5779	{
5780	# if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64)
5781	_InterlockedAnd64((__int64 volatile RT_FAR *)pu64, u64);
5782
5783	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
5784	__asm__ __volatile__("lock; andq %1, %0\n\t"
5785	: "=m" (*pu64)
5786	: "r" (u64)
5787	, "m" (*pu64)
5788	: "cc");
5789
5790	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5791	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5792	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5793	uint64_t u64Spill;
5794	__asm__ __volatile__("Lstart_ASMAtomicAndU64_%=:\n\t"
5795	"ldclral %[fBitsToClear], %[uSpill], %[pMem]\n\t"
5796	: [pMem] "+Q" (*pu64)
5797	, [uSpill] "=&r" (u64Spill)
5798	: [fBitsToClear] "r" (~u64)
5799	: );
5800	# else
5801	__asm__ __volatile__("Lstart_ASMAtomicAndU64_%=:\n\t"
5802	RTASM_ARM_DMB_SY
5803	"stclr %[fBitsToClear], %[pMem]\n\t"
5804	: [pMem] "+Q" (*pu64)
5805	: [fBitsToClear] "r" (~u64)
5806	: );
5807	# endif
5808	# else
5809	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicAndU64, pu64, DMB_SY,
5810	"and %[uNew], %[uNew], %[uVal]\n\t"
5811	,
5812	"and %[uNew], %[uNew], %[uVal]\n\t"
5813	"and %H[uNew], %H[uNew], %H[uVal]\n\t",
5814	[uVal] "r" (u64));
5815	# endif
5816
5817	# else
5818	for (;;)
5819	{
5820	uint64_t u64Old = ASMAtomicUoReadU64(pu64);
5821	uint64_t u64New = u64Old & u64;
5822	if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old))
5823	break;
5824	ASMNopPause();
5825	}
5826	# endif
5827	}
5828	#endif
5829
5830
5831	/**
5832	* Atomically And a signed 64-bit value, ordered.
5833	*
5834	* @param pi64 Pointer to the pointer variable to AND i64 with.
5835	* @param i64 The value to AND *pi64 with.
5836	*
5837	* @remarks x86: Requires a Pentium or later.
5838	*/
5839	DECLINLINE(void) ASMAtomicAndS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
5840	{
5841	ASMAtomicAndU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64);
5842	}
5843
5844
5845	/**
5846	* Atomically XOR an unsigned 32-bit value and a memory location, ordered.
5847	*
5848	* @param pu32 Pointer to the variable to XOR @a u32 with.
5849	* @param u32 The value to XOR @a *pu32 with.
5850	*
5851	* @remarks x86: Requires a 386 or later.
5852	*/
5853	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
5854	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
5855	#else
5856	DECLINLINE(void) ASMAtomicXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5857	{
5858	# if RT_INLINE_ASM_USES_INTRIN
5859	_InterlockedXor((long volatile RT_FAR *)pu32, u32);
5860
5861	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5862	# if RT_INLINE_ASM_GNU_STYLE
5863	__asm__ __volatile__("lock; xorl %1, %0\n\t"
5864	: "=m" (*pu32)
5865	: "ir" (u32)
5866	, "m" (*pu32)
5867	: "cc");
5868	# else
5869	__asm
5870	{
5871	mov eax, [u32]
5872	# ifdef RT_ARCH_AMD64
5873	mov rdx, [pu32]
5874	lock xor [rdx], eax
5875	# else
5876	mov edx, [pu32]
5877	lock xor [edx], eax
5878	# endif
5879	}
5880	# endif
5881
5882	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5883	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5884	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5885	uint32_t u32Spill;
5886	__asm__ __volatile__("Lstart_ASMAtomicXorU32_%=:\n\t"
5887	"ldeoral %w[fBitMask], %w[uSpill], %[pMem]\n\t"
5888	: [pMem] "+Q" (*pu32)
5889	, [uSpill] "=&r" (u32Spill)
5890	: [fBitMask] "r" (u32)
5891	: );
5892	# else
5893	__asm__ __volatile__("Lstart_ASMAtomicXorU32_%=:\n\t"
5894	RTASM_ARM_DMB_SY
5895	"steor %w[fBitMask], %[pMem]\n\t"
5896	: [pMem] "+Q" (*pu32)
5897	: [fBitMask] "r" (u32)
5898	: );
5899	# endif
5900	# else
5901	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicXor32, pu32, DMB_SY,
5902	"eor %w[uNew], %w[uNew], %w[uVal]\n\t",
5903	"eor %[uNew], %[uNew], %[uVal]\n\t",
5904	[uVal] "r" (u32));
5905	# endif
5906
5907	# else
5908	# error "Port me"
5909	# endif
5910	}
5911	#endif
5912
5913
5914	/**
5915	* Atomically XOR an unsigned 32-bit value and a memory location, ordered,
5916	* extended version (for bitmaps).
5917	*
5918	* @returns Old value.
5919	* @param pu32 Pointer to the variable to XOR @a u32 with.
5920	* @param u32 The value to XOR @a *pu32 with.
5921	*/
5922	DECLINLINE(uint32_t) ASMAtomicXorExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5923	{
5924	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
5925	# if defined(RTASM_ARM64_USE_FEAT_LSE)
5926	uint32_t u32OldRet;
5927	__asm__ __volatile__("Lstart_ASMAtomicXorExU32_%=:\n\t"
5928	# if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB)
5929	"ldeoral %w[fBitMask], %w[uOldRet], %[pMem]\n\t"
5930	# else
5931	RTASM_ARM_DMB_SY
5932	"ldeor %w[fBitMask], %w[uOldRet], %[pMem]\n\t"
5933	# endif
5934	: [pMem] "+Q" (*pu32)
5935	, [uOldRet] "=&r" (u32OldRet)
5936	: [fBitMask] "r" (u32)
5937	: );
5938	# else
5939	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicXorEx32, pu32, DMB_SY,
5940	"eor %w[uNew], %w[uOld], %w[uVal]\n\t",
5941	"eor %[uNew], %[uOld], %[uVal]\n\t",
5942	[uVal] "r" (u32));
5943	# endif
5944	return u32OldRet;
5945
5946	#else
5947	uint32_t u32RetOld = ASMAtomicUoReadU32(pu32);
5948	uint32_t u32New;
5949	do
5950	u32New = u32RetOld ^ u32;
5951	while (!ASMAtomicCmpXchgExU32(pu32, u32New, u32RetOld, &u32RetOld));
5952	return u32RetOld;
5953	#endif
5954	}
5955
5956
5957	/**
5958	* Atomically XOR a signed 32-bit value, ordered.
5959	*
5960	* @param pi32 Pointer to the variable to XOR i32 with.
5961	* @param i32 The value to XOR *pi32 with.
5962	*
5963	* @remarks x86: Requires a 386 or later.
5964	*/
5965	DECLINLINE(void) ASMAtomicXorS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
5966	{
5967	ASMAtomicXorU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
5968	}
5969
5970
5971	/**
5972	* Atomically OR an unsigned 32-bit value, unordered but interrupt safe.
5973	*
5974	* @param pu32 Pointer to the pointer variable to OR u32 with.
5975	* @param u32 The value to OR *pu32 with.
5976	*
5977	* @remarks x86: Requires a 386 or later.
5978	*/
5979	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
5980	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicUoOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
5981	#else
5982	DECLINLINE(void) ASMAtomicUoOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
5983	{
5984	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
5985	# if RT_INLINE_ASM_GNU_STYLE
5986	__asm__ __volatile__("orl %1, %0\n\t"
5987	: "=m" (*pu32)
5988	: "ir" (u32)
5989	, "m" (*pu32)
5990	: "cc");
5991	# else
5992	__asm
5993	{
5994	mov eax, [u32]
5995	# ifdef RT_ARCH_AMD64
5996	mov rdx, [pu32]
5997	or [rdx], eax
5998	# else
5999	mov edx, [pu32]
6000	or [edx], eax
6001	# endif
6002	}
6003	# endif
6004
6005	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6006	/* M1 benchmark: stset=1974 vs non-lse=6271 */
6007	# if defined(RTASM_ARM64_USE_FEAT_LSE)
6008	__asm__ __volatile__("Lstart_ASMAtomicUoOrU32_%=:\n\t"
6009	"stset %w[fBitsToSet], %[pMem]\n\t"
6010	: [pMem] "+Q" (*pu32)
6011	: [fBitsToSet] "r" (u32)
6012	: );
6013	# else
6014	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoOrU32, pu32, NO_BARRIER,
6015	"orr %w[uNew], %w[uNew], %w[uVal]\n\t",
6016	"orr %[uNew], %[uNew], %[uVal]\n\t",
6017	[uVal] "r" (u32));
6018	# endif
6019
6020	# else
6021	# error "Port me"
6022	# endif
6023	}
6024	#endif
6025
6026
6027	/**
6028	* Atomically OR an unsigned 32-bit value, unordered but interrupt safe,
6029	* extended version (for bitmap fallback).
6030	*
6031	* @returns Old value.
6032	* @param pu32 Pointer to the variable to OR @a u32 with.
6033	* @param u32 The value to OR @a *pu32 with.
6034	*/
6035	DECLINLINE(uint32_t) ASMAtomicUoOrExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
6036	{
6037	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6038	# if defined(RTASM_ARM64_USE_FEAT_LSE)
6039	uint32_t u32OldRet;
6040	__asm__ __volatile__("Lstart_ASMAtomicOrExU32_%=:\n\t"
6041	"ldset %w[fBitsToSet], %w[uOldRet], %[pMem]\n\t"
6042	: [pMem] "+Q" (*pu32)
6043	, [uOldRet] "=&r" (u32OldRet)
6044	: [fBitsToSet] "r" (u32)
6045	: );
6046	# else
6047	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicUoOrExU32, pu32, NO_BARRIER,
6048	"orr %w[uNew], %w[uOld], %w[uVal]\n\t",
6049	"orr %[uNew], %[uOld], %[uVal]\n\t",
6050	[uVal] "r" (u32));
6051	# endif
6052	return u32OldRet;
6053
6054	#else
6055	return ASMAtomicOrExU32(pu32, u32); /* (we have no unordered cmpxchg primitive atm.) */
6056	#endif
6057	}
6058
6059
6060	/**
6061	* Atomically OR a signed 32-bit value, unordered.
6062	*
6063	* @param pi32 Pointer to the pointer variable to OR u32 with.
6064	* @param i32 The value to OR *pu32 with.
6065	*
6066	* @remarks x86: Requires a 386 or later.
6067	*/
6068	DECLINLINE(void) ASMAtomicUoOrS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
6069	{
6070	ASMAtomicUoOrU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
6071	}
6072
6073
6074	/**
6075	* Atomically OR an unsigned 64-bit value, unordered.
6076	*
6077	* @param pu64 Pointer to the pointer variable to OR u64 with.
6078	* @param u64 The value to OR *pu64 with.
6079	*
6080	* @remarks x86: Requires a Pentium or later.
6081	*/
6082	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
6083	DECLASM(void) ASMAtomicUoOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
6084	#else
6085	DECLINLINE(void) ASMAtomicUoOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
6086	{
6087	# if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
6088	__asm__ __volatile__("orq %1, %q0\n\t"
6089	: "=m" (*pu64)
6090	: "r" (u64)
6091	, "m" (*pu64)
6092	: "cc");
6093
6094	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6095	# if defined(RTASM_ARM64_USE_FEAT_LSE)
6096	__asm__ __volatile__("Lstart_ASMAtomicUoOrU64_%=:\n\t"
6097	"stset %[fBitsToSet], %[pMem]\n\t"
6098	: [pMem] "+Q" (*pu64)
6099	: [fBitsToSet] "r" (u64)
6100	: );
6101	# else
6102	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicUoOrU64, pu64, NO_BARRIER,
6103	"orr %[uNew], %[uNew], %[uVal]\n\t"
6104	,
6105	"orr %[uNew], %[uNew], %[uVal]\n\t"
6106	"orr %H[uNew], %H[uNew], %H[uVal]\n\t",
6107	[uVal] "r" (u64));
6108	# endif
6109
6110	# else
6111	for (;;)
6112	{
6113	uint64_t u64Old = ASMAtomicUoReadU64(pu64);
6114	uint64_t u64New = u64Old \| u64;
6115	if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old))
6116	break;
6117	ASMNopPause();
6118	}
6119	# endif
6120	}
6121	#endif
6122
6123
6124	/**
6125	* Atomically Or a signed 64-bit value, unordered.
6126	*
6127	* @param pi64 Pointer to the pointer variable to OR u64 with.
6128	* @param i64 The value to OR *pu64 with.
6129	*
6130	* @remarks x86: Requires a Pentium or later.
6131	*/
6132	DECLINLINE(void) ASMAtomicUoOrS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
6133	{
6134	ASMAtomicUoOrU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64);
6135	}
6136
6137
6138	/**
6139	* Atomically And an unsigned 32-bit value, unordered.
6140	*
6141	* @param pu32 Pointer to the pointer variable to AND u32 with.
6142	* @param u32 The value to AND *pu32 with.
6143	*
6144	* @remarks x86: Requires a 386 or later.
6145	*/
6146	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
6147	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicUoAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
6148	#else
6149	DECLINLINE(void) ASMAtomicUoAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
6150	{
6151	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6152	# if RT_INLINE_ASM_GNU_STYLE
6153	__asm__ __volatile__("andl %1, %0\n\t"
6154	: "=m" (*pu32)
6155	: "ir" (u32)
6156	, "m" (*pu32)
6157	: "cc");
6158	# else
6159	__asm
6160	{
6161	mov eax, [u32]
6162	# ifdef RT_ARCH_AMD64
6163	mov rdx, [pu32]
6164	and [rdx], eax
6165	# else
6166	mov edx, [pu32]
6167	and [edx], eax
6168	# endif
6169	}
6170	# endif
6171
6172	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6173	/* M1 benchmark: stclr=1884 vs non-lse=6299 (ps/call) */
6174	# if defined(RTASM_ARM64_USE_FEAT_LSE)
6175	__asm__ __volatile__("Lstart_ASMAtomicUoAndU32_%=:\n\t"
6176	"stclr %w[fBitsToClear], %[pMem]\n\t"
6177	: [pMem] "+Q" (*pu32)
6178	: [fBitsToClear] "r" (~u32)
6179	: );
6180	# else
6181	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoAnd32, pu32, NO_BARRIER,
6182	"and %w[uNew], %w[uNew], %w[uVal]\n\t",
6183	"and %[uNew], %[uNew], %[uVal]\n\t",
6184	[uVal] "r" (u32));
6185	# endif
6186
6187	# else
6188	# error "Port me"
6189	# endif
6190	}
6191	#endif
6192
6193
6194	/**
6195	* Atomically AND an unsigned 32-bit value, unordered, extended version (for
6196	* bitmap fallback).
6197	*
6198	* @returns Old value.
6199	* @param pu32 Pointer to the pointer to AND @a u32 with.
6200	* @param u32 The value to AND @a *pu32 with.
6201	*/
6202	DECLINLINE(uint32_t) ASMAtomicUoAndExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
6203	{
6204	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6205	# if defined(RTASM_ARM64_USE_FEAT_LSE)
6206	uint32_t u32OldRet;
6207	__asm__ __volatile__("Lstart_ASMAtomicAndExU32_%=:\n\t"
6208	"ldclr %w[fBitsToClear], %w[uOldRet], %[pMem]\n\t"
6209	: [pMem] "+Q" (*pu32)
6210	, [uOldRet] "=&r" (u32OldRet)
6211	: [fBitsToClear] "r" (~u32)
6212	: );
6213	# else
6214	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicUoAndEx32, pu32, NO_BARRIER,
6215	"and %w[uNew], %w[uOld], %w[uVal]\n\t",
6216	"and %[uNew], %[uOld], %[uVal]\n\t",
6217	[uVal] "r" (u32));
6218	# endif
6219	return u32OldRet;
6220
6221	#else
6222	return ASMAtomicAndExU32(pu32, u32); /* (we have no unordered cmpxchg primitive atm.) */
6223	#endif
6224	}
6225
6226
6227	/**
6228	* Atomically And a signed 32-bit value, unordered.
6229	*
6230	* @param pi32 Pointer to the pointer variable to AND i32 with.
6231	* @param i32 The value to AND *pi32 with.
6232	*
6233	* @remarks x86: Requires a 386 or later.
6234	*/
6235	DECLINLINE(void) ASMAtomicUoAndS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
6236	{
6237	ASMAtomicUoAndU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
6238	}
6239
6240
6241	/**
6242	* Atomically And an unsigned 64-bit value, unordered.
6243	*
6244	* @param pu64 Pointer to the pointer variable to AND u64 with.
6245	* @param u64 The value to AND *pu64 with.
6246	*
6247	* @remarks x86: Requires a Pentium or later.
6248	*/
6249	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
6250	DECLASM(void) ASMAtomicUoAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO;
6251	#else
6252	DECLINLINE(void) ASMAtomicUoAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF
6253	{
6254	# if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
6255	__asm__ __volatile__("andq %1, %0\n\t"
6256	: "=m" (*pu64)
6257	: "r" (u64)
6258	, "m" (*pu64)
6259	: "cc");
6260
6261	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6262	# if defined(RTASM_ARM64_USE_FEAT_LSE)
6263	__asm__ __volatile__("Lstart_ASMAtomicUoAndU64_%=:\n\t"
6264	"stclr %[fBitsToClear], %[pMem]\n\t"
6265	: [pMem] "+Q" (*pu64)
6266	: [fBitsToClear] "r" (~u64)
6267	: );
6268	# else
6269	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicUoAndU64, pu64, NO_BARRIER,
6270	"and %[uNew], %[uNew], %[uVal]\n\t"
6271	,
6272	"and %[uNew], %[uNew], %[uVal]\n\t"
6273	"and %H[uNew], %H[uNew], %H[uVal]\n\t",
6274	[uVal] "r" (u64));
6275	# endif
6276
6277	# else
6278	for (;;)
6279	{
6280	uint64_t u64Old = ASMAtomicUoReadU64(pu64);
6281	uint64_t u64New = u64Old & u64;
6282	if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old))
6283	break;
6284	ASMNopPause();
6285	}
6286	# endif
6287	}
6288	#endif
6289
6290
6291	/**
6292	* Atomically And a signed 64-bit value, unordered.
6293	*
6294	* @param pi64 Pointer to the pointer variable to AND i64 with.
6295	* @param i64 The value to AND *pi64 with.
6296	*
6297	* @remarks x86: Requires a Pentium or later.
6298	*/
6299	DECLINLINE(void) ASMAtomicUoAndS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF
6300	{
6301	ASMAtomicUoAndU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64);
6302	}
6303
6304
6305	/**
6306	* Atomically XOR an unsigned 32-bit value, unordered but interrupt safe.
6307	*
6308	* @param pu32 Pointer to the variable to XOR @a u32 with.
6309	* @param u32 The value to OR @a *pu32 with.
6310	*
6311	* @remarks x86: Requires a 386 or later.
6312	*/
6313	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
6314	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicUoXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO;
6315	#else
6316	DECLINLINE(void) ASMAtomicUoXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
6317	{
6318	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6319	# if RT_INLINE_ASM_GNU_STYLE
6320	__asm__ __volatile__("xorl %1, %0\n\t"
6321	: "=m" (*pu32)
6322	: "ir" (u32)
6323	, "m" (*pu32)
6324	: "cc");
6325	# else
6326	__asm
6327	{
6328	mov eax, [u32]
6329	# ifdef RT_ARCH_AMD64
6330	mov rdx, [pu32]
6331	xor [rdx], eax
6332	# else
6333	mov edx, [pu32]
6334	xor [edx], eax
6335	# endif
6336	}
6337	# endif
6338
6339	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6340	# if defined(RTASM_ARM64_USE_FEAT_LSE)
6341	__asm__ __volatile__("Lstart_ASMAtomicUoXorU32_%=:\n\t"
6342	"steor %w[fBitMask], %[pMem]\n\t"
6343	: [pMem] "+Q" (*pu32)
6344	: [fBitMask] "r" (u32)
6345	: );
6346	# else
6347	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoXorU32, pu32, NO_BARRIER,
6348	"eor %w[uNew], %w[uNew], %w[uVal]\n\t",
6349	"eor %[uNew], %[uNew], %[uVal]\n\t",
6350	[uVal] "r" (u32));
6351	# endif
6352
6353	# else
6354	# error "Port me"
6355	# endif
6356	}
6357	#endif
6358
6359
6360	/**
6361	* Atomically XOR an unsigned 32-bit value, unordered but interrupt safe,
6362	* extended version (for bitmap fallback).
6363	*
6364	* @returns Old value.
6365	* @param pu32 Pointer to the variable to XOR @a u32 with.
6366	* @param u32 The value to OR @a *pu32 with.
6367	*/
6368	DECLINLINE(uint32_t) ASMAtomicUoXorExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF
6369	{
6370	#if defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6371	# if defined(RTASM_ARM64_USE_FEAT_LSE)
6372	uint32_t u32OldRet;
6373	__asm__ __volatile__("Lstart_ASMAtomicUoXorExU32_%=:\n\t"
6374	"ldeor %w[fBitMask], %w[uOldRet], %[pMem]\n\t"
6375	: [pMem] "+Q" (*pu32)
6376	, [uOldRet] "=&r" (u32OldRet)
6377	: [fBitMask] "r" (u32)
6378	: );
6379	# else
6380	RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicUoXorExU32, pu32, NO_BARRIER,
6381	"eor %w[uNew], %w[uOld], %w[uVal]\n\t",
6382	"eor %[uNew], %[uOld], %[uVal]\n\t",
6383	[uVal] "r" (u32));
6384	# endif
6385	return u32OldRet;
6386
6387	#else
6388	return ASMAtomicXorExU32(pu32, u32); /* (we have no unordered cmpxchg primitive atm.) */
6389	#endif
6390	}
6391
6392
6393	/**
6394	* Atomically XOR a signed 32-bit value, unordered.
6395	*
6396	* @param pi32 Pointer to the variable to XOR @a u32 with.
6397	* @param i32 The value to XOR @a *pu32 with.
6398	*
6399	* @remarks x86: Requires a 386 or later.
6400	*/
6401	DECLINLINE(void) ASMAtomicUoXorS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF
6402	{
6403	ASMAtomicUoXorU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32);
6404	}
6405
6406
6407	/**
6408	* Atomically increment an unsigned 32-bit value, unordered.
6409	*
6410	* @returns the new value.
6411	* @param pu32 Pointer to the variable to increment.
6412	*
6413	* @remarks x86: Requires a 486 or later.
6414	*/
6415	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
6416	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicUoIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO;
6417	#else
6418	DECLINLINE(uint32_t) ASMAtomicUoIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF
6419	{
6420	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6421	uint32_t u32;
6422	# if RT_INLINE_ASM_GNU_STYLE
6423	__asm__ __volatile__("xaddl %0, %1\n\t"
6424	: "=r" (u32)
6425	, "=m" (*pu32)
6426	: "0" (1)
6427	, "m" (*pu32)
6428	: "memory" /** @todo why 'memory'? */
6429	, "cc");
6430	return u32 + 1;
6431	# else
6432	__asm
6433	{
6434	mov eax, 1
6435	# ifdef RT_ARCH_AMD64
6436	mov rdx, [pu32]
6437	xadd [rdx], eax
6438	# else
6439	mov edx, [pu32]
6440	xadd [edx], eax
6441	# endif
6442	mov u32, eax
6443	}
6444	return u32 + 1;
6445	# endif
6446
6447	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6448	/* M1 benchmark: ldadd=2031 vs non-lse=6301 (ps/call) */
6449	# if defined(RTASM_ARM64_USE_FEAT_LSE)
6450	uint32_t u32NewRet;
6451	__asm__ __volatile__("Lstart_ASMAtomicUoIncU32_%=:\n\t"
6452	"ldadd %w[uAddend], %w[uNewRet], %[pMem]\n\t"
6453	"add %w[uNewRet], %w[uNewRet], #1\n\t"
6454	: [pMem] "+Q" (*pu32)
6455	, [uNewRet] "=&r" (u32NewRet)
6456	: [uAddend] "r" ((uint32_t)1)
6457	: );
6458	# else
6459	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoIncU32, pu32, NO_BARRIER,
6460	"add %w[uNew], %w[uNew], #1\n\t",
6461	"add %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */,
6462	"X" (0) /* dummy */);
6463	# endif
6464	return u32NewRet;
6465
6466	# else
6467	# error "Port me"
6468	# endif
6469	}
6470	#endif
6471
6472
6473	/**
6474	* Atomically decrement an unsigned 32-bit value, unordered.
6475	*
6476	* @returns the new value.
6477	* @param pu32 Pointer to the variable to decrement.
6478	*
6479	* @remarks x86: Requires a 486 or later.
6480	*/
6481	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
6482	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicUoDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO;
6483	#else
6484	DECLINLINE(uint32_t) ASMAtomicUoDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF
6485	{
6486	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6487	uint32_t u32;
6488	# if RT_INLINE_ASM_GNU_STYLE
6489	__asm__ __volatile__("lock; xaddl %0, %1\n\t"
6490	: "=r" (u32)
6491	, "=m" (*pu32)
6492	: "0" (-1)
6493	, "m" (*pu32)
6494	: "memory"
6495	, "cc");
6496	return u32 - 1;
6497	# else
6498	__asm
6499	{
6500	mov eax, -1
6501	# ifdef RT_ARCH_AMD64
6502	mov rdx, [pu32]
6503	xadd [rdx], eax
6504	# else
6505	mov edx, [pu32]
6506	xadd [edx], eax
6507	# endif
6508	mov u32, eax
6509	}
6510	return u32 - 1;
6511	# endif
6512
6513	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6514	/* M1 benchmark: ldadd=2101 vs non-lse=6268 (ps/call) */
6515	# if defined(RTASM_ARM64_USE_FEAT_LSE)
6516	uint32_t u32NewRet;
6517	__asm__ __volatile__("Lstart_ASMAtomicUoDecU32_%=:\n\t"
6518	"ldadd %w[uAddend], %w[uNewRet], %[pMem]\n\t"
6519	"sub %w[uNewRet], %w[uNewRet], #1\n\t"
6520	: [pMem] "+Q" (*pu32)
6521	, [uNewRet] "=&r" (u32NewRet)
6522	: [uAddend] "r" (~(uint32_t)0)
6523	: );
6524	# else
6525	RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoDecU32, pu32, NO_BARRIER,
6526	"sub %w[uNew], %w[uNew], #1\n\t",
6527	"sub %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */,
6528	"X" (0) /* dummy */);
6529	# endif
6530	return u32NewRet;
6531
6532	# else
6533	# error "Port me"
6534	# endif
6535	}
6536	#endif
6537
6538	/** @todo Move ASMByteSwapU16, ASMByteSwapU32 and ASMByteSwapU64 in their own
6539	* header as it's a common reason for including asm.h. */
6540
6541
6542	/**
6543	* Reverse the byte order of the given 16-bit integer.
6544	*
6545	* @returns Revert
6546	* @param u16 16-bit integer value.
6547	*/
6548	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6549	RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMByteSwapU16(uint16_t u16) RT_NOTHROW_PROTO;
6550	#else
6551	DECLINLINE(uint16_t) ASMByteSwapU16(uint16_t u16) RT_NOTHROW_DEF
6552	{
6553	# if RT_INLINE_ASM_USES_INTRIN
6554	return _byteswap_ushort(u16);
6555
6556	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6557	# if RT_INLINE_ASM_GNU_STYLE
6558	__asm__ ("rorw $8, %0" : "=r" (u16) : "0" (u16) : "cc");
6559	# else
6560	_asm
6561	{
6562	mov ax, [u16]
6563	ror ax, 8
6564	mov [u16], ax
6565	}
6566	# endif
6567	return u16;
6568
6569	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
6570	uint32_t u32Ret;
6571	__asm__ __volatile__(
6572	# if defined(RT_ARCH_ARM64)
6573	"rev16 %w[uRet], %w[uVal]\n\t"
6574	# else
6575	"rev16 %[uRet], %[uVal]\n\t"
6576	# endif
6577	: [uRet] "=r" (u32Ret)
6578	: [uVal] "r" (u16));
6579	return (uint16_t)u32Ret;
6580
6581	# else
6582	# error "Port me"
6583	# endif
6584	}
6585	#endif
6586
6587
6588	/**
6589	* Reverse the byte order of the given 32-bit integer.
6590	*
6591	* @returns Revert
6592	* @param u32 32-bit integer value.
6593	*/
6594	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6595	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMByteSwapU32(uint32_t u32) RT_NOTHROW_PROTO;
6596	#else
6597	DECLINLINE(uint32_t) ASMByteSwapU32(uint32_t u32) RT_NOTHROW_DEF
6598	{
6599	# if RT_INLINE_ASM_USES_INTRIN
6600	return _byteswap_ulong(u32);
6601
6602	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6603	# if RT_INLINE_ASM_GNU_STYLE
6604	__asm__ ("bswapl %0" : "=r" (u32) : "0" (u32));
6605	# else
6606	_asm
6607	{
6608	mov eax, [u32]
6609	bswap eax
6610	mov [u32], eax
6611	}
6612	# endif
6613	return u32;
6614
6615	# elif defined(RT_ARCH_ARM64)
6616	uint64_t u64Ret;
6617	__asm__ __volatile__("rev32 %[uRet], %[uVal]\n\t"
6618	: [uRet] "=r" (u64Ret)
6619	: [uVal] "r" ((uint64_t)u32));
6620	return (uint32_t)u64Ret;
6621
6622	# elif defined(RT_ARCH_ARM32)
6623	__asm__ __volatile__("rev %[uRet], %[uVal]\n\t"
6624	: [uRet] "=r" (u32)
6625	: [uVal] "[uRet]" (u32));
6626	return u32;
6627
6628	# else
6629	# error "Port me"
6630	# endif
6631	}
6632	#endif
6633
6634
6635	/**
6636	* Reverse the byte order of the given 64-bit integer.
6637	*
6638	* @returns Revert
6639	* @param u64 64-bit integer value.
6640	*/
6641	DECLINLINE(uint64_t) ASMByteSwapU64(uint64_t u64) RT_NOTHROW_DEF
6642	{
6643	#if defined(RT_ARCH_AMD64) && RT_INLINE_ASM_USES_INTRIN
6644	return _byteswap_uint64(u64);
6645
6646	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
6647	__asm__ ("bswapq %0" : "=r" (u64) : "0" (u64));
6648	return u64;
6649
6650	# elif defined(RT_ARCH_ARM64)
6651	__asm__ __volatile__("rev %[uRet], %[uVal]\n\t"
6652	: [uRet] "=r" (u64)
6653	: [uVal] "[uRet]" (u64));
6654	return u64;
6655
6656	#else
6657	return (uint64_t)ASMByteSwapU32((uint32_t)u64) << 32
6658	\| (uint64_t)ASMByteSwapU32((uint32_t)(u64 >> 32));
6659	#endif
6660	}
6661
6662
6663
6664	/** @defgroup grp_inline_bits Bitmap Operations
6665	*
6666	* @todo Move these into a separate header, with standard IPRT prefix
6667	* (RTBitmapXxx). Move the more complex (searched) stuff into C source
6668	* files.
6669	*
6670	* @{
6671	*/
6672
6673
6674	/**
6675	* Sets a bit in a bitmap.
6676	*
6677	* @param pvBitmap Pointer to the bitmap (little endian). This should be
6678	* 32-bit aligned.
6679	* @param iBit The bit to set.
6680	*
6681	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
6682	* However, doing so will yield better performance as well as avoiding
6683	* traps accessing the last bits in the bitmap.
6684	*/
6685	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6686	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6687	#else
6688	DECLINLINE(void) ASMBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6689	{
6690	# if RT_INLINE_ASM_USES_INTRIN
6691	_bittestandset((long RT_FAR *)pvBitmap, iBit);
6692
6693	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6694	# if RT_INLINE_ASM_GNU_STYLE
6695	__asm__ __volatile__("btsl %1, %0"
6696	: "=m" ((volatile long RT_FAR )pvBitmap)
6697	: "Ir" (iBit)
6698	, "m" ((volatile long RT_FAR )pvBitmap)
6699	: "memory"
6700	, "cc");
6701	# else
6702	__asm
6703	{
6704	# ifdef RT_ARCH_AMD64
6705	mov rax, [pvBitmap]
6706	mov edx, [iBit]
6707	bts [rax], edx
6708	# else
6709	mov eax, [pvBitmap]
6710	mov edx, [iBit]
6711	bts [eax], edx
6712	# endif
6713	}
6714	# endif
6715
6716	# else
6717	int32_t offBitmap = iBit / 32;
6718	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
6719	ASMAtomicUoOrU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31)));
6720	# endif
6721	}
6722	#endif
6723
6724
6725	/**
6726	* Atomically sets a bit in a bitmap, ordered.
6727	*
6728	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
6729	* aligned, otherwise the memory access isn't atomic!
6730	* @param iBit The bit to set.
6731	*
6732	* @remarks x86: Requires a 386 or later.
6733	*/
6734	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6735	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6736	#else
6737	DECLINLINE(void) ASMAtomicBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6738	{
6739	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
6740	# if RT_INLINE_ASM_USES_INTRIN
6741	_interlockedbittestandset((long RT_FAR *)pvBitmap, iBit);
6742	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6743	# if RT_INLINE_ASM_GNU_STYLE
6744	__asm__ __volatile__("lock; btsl %1, %0"
6745	: "=m" ((volatile long )pvBitmap)
6746	: "Ir" (iBit)
6747	, "m" ((volatile long )pvBitmap)
6748	: "memory"
6749	, "cc");
6750	# else
6751	__asm
6752	{
6753	# ifdef RT_ARCH_AMD64
6754	mov rax, [pvBitmap]
6755	mov edx, [iBit]
6756	lock bts [rax], edx
6757	# else
6758	mov eax, [pvBitmap]
6759	mov edx, [iBit]
6760	lock bts [eax], edx
6761	# endif
6762	}
6763	# endif
6764
6765	# else
6766	ASMAtomicOrU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(RT_BIT_32(iBit & 31)));
6767	# endif
6768	}
6769	#endif
6770
6771
6772	/**
6773	* Clears a bit in a bitmap.
6774	*
6775	* @param pvBitmap Pointer to the bitmap (little endian).
6776	* @param iBit The bit to clear.
6777	*
6778	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
6779	* However, doing so will yield better performance as well as avoiding
6780	* traps accessing the last bits in the bitmap.
6781	*/
6782	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6783	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6784	#else
6785	DECLINLINE(void) ASMBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6786	{
6787	# if RT_INLINE_ASM_USES_INTRIN
6788	_bittestandreset((long RT_FAR *)pvBitmap, iBit);
6789
6790	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6791	# if RT_INLINE_ASM_GNU_STYLE
6792	__asm__ __volatile__("btrl %1, %0"
6793	: "=m" ((volatile long RT_FAR )pvBitmap)
6794	: "Ir" (iBit)
6795	, "m" ((volatile long RT_FAR )pvBitmap)
6796	: "memory"
6797	, "cc");
6798	# else
6799	__asm
6800	{
6801	# ifdef RT_ARCH_AMD64
6802	mov rax, [pvBitmap]
6803	mov edx, [iBit]
6804	btr [rax], edx
6805	# else
6806	mov eax, [pvBitmap]
6807	mov edx, [iBit]
6808	btr [eax], edx
6809	# endif
6810	}
6811	# endif
6812
6813	# else
6814	int32_t offBitmap = iBit / 32;
6815	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
6816	ASMAtomicUoAndU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(~RT_BIT_32(iBit & 31)));
6817	# endif
6818	}
6819	#endif
6820
6821
6822	/**
6823	* Atomically clears a bit in a bitmap, ordered.
6824	*
6825	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
6826	* aligned, otherwise the memory access isn't atomic!
6827	* @param iBit The bit to toggle set.
6828	*
6829	* @remarks No memory barrier, take care on smp.
6830	* @remarks x86: Requires a 386 or later.
6831	*/
6832	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
6833	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6834	#else
6835	DECLINLINE(void) ASMAtomicBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6836	{
6837	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
6838	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6839	# if RT_INLINE_ASM_GNU_STYLE
6840	__asm__ __volatile__("lock; btrl %1, %0"
6841	: "=m" ((volatile long RT_FAR )pvBitmap)
6842	: "Ir" (iBit)
6843	, "m" ((volatile long RT_FAR )pvBitmap)
6844	: "memory"
6845	, "cc");
6846	# else
6847	__asm
6848	{
6849	# ifdef RT_ARCH_AMD64
6850	mov rax, [pvBitmap]
6851	mov edx, [iBit]
6852	lock btr [rax], edx
6853	# else
6854	mov eax, [pvBitmap]
6855	mov edx, [iBit]
6856	lock btr [eax], edx
6857	# endif
6858	}
6859	# endif
6860	# else
6861	ASMAtomicAndU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(~RT_BIT_32(iBit & 31)));
6862	# endif
6863	}
6864	#endif
6865
6866
6867	/**
6868	* Toggles a bit in a bitmap.
6869	*
6870	* @param pvBitmap Pointer to the bitmap (little endian).
6871	* @param iBit The bit to toggle.
6872	*
6873	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
6874	* However, doing so will yield better performance as well as avoiding
6875	* traps accessing the last bits in the bitmap.
6876	*/
6877	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6878	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6879	#else
6880	DECLINLINE(void) ASMBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6881	{
6882	# if RT_INLINE_ASM_USES_INTRIN
6883	_bittestandcomplement((long RT_FAR *)pvBitmap, iBit);
6884	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6885	# if RT_INLINE_ASM_GNU_STYLE
6886	__asm__ __volatile__("btcl %1, %0"
6887	: "=m" ((volatile long )pvBitmap)
6888	: "Ir" (iBit)
6889	, "m" ((volatile long )pvBitmap)
6890	: "memory"
6891	, "cc");
6892	# else
6893	__asm
6894	{
6895	# ifdef RT_ARCH_AMD64
6896	mov rax, [pvBitmap]
6897	mov edx, [iBit]
6898	btc [rax], edx
6899	# else
6900	mov eax, [pvBitmap]
6901	mov edx, [iBit]
6902	btc [eax], edx
6903	# endif
6904	}
6905	# endif
6906	# else
6907	int32_t offBitmap = iBit / 32;
6908	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
6909	ASMAtomicUoXorU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31)));
6910	# endif
6911	}
6912	#endif
6913
6914
6915	/**
6916	* Atomically toggles a bit in a bitmap, ordered.
6917	*
6918	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
6919	* aligned, otherwise the memory access isn't atomic!
6920	* @param iBit The bit to test and set.
6921	*
6922	* @remarks x86: Requires a 386 or later.
6923	*/
6924	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
6925	RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6926	#else
6927	DECLINLINE(void) ASMAtomicBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6928	{
6929	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
6930	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6931	# if RT_INLINE_ASM_GNU_STYLE
6932	__asm__ __volatile__("lock; btcl %1, %0"
6933	: "=m" ((volatile long RT_FAR )pvBitmap)
6934	: "Ir" (iBit)
6935	, "m" ((volatile long RT_FAR )pvBitmap)
6936	: "memory"
6937	, "cc");
6938	# else
6939	__asm
6940	{
6941	# ifdef RT_ARCH_AMD64
6942	mov rax, [pvBitmap]
6943	mov edx, [iBit]
6944	lock btc [rax], edx
6945	# else
6946	mov eax, [pvBitmap]
6947	mov edx, [iBit]
6948	lock btc [eax], edx
6949	# endif
6950	}
6951	# endif
6952	# else
6953	ASMAtomicXorU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(RT_BIT_32(iBit & 31)));
6954	# endif
6955	}
6956	#endif
6957
6958
6959	/**
6960	* Tests and sets a bit in a bitmap.
6961	*
6962	* @returns true if the bit was set.
6963	* @returns false if the bit was clear.
6964	*
6965	* @param pvBitmap Pointer to the bitmap (little endian).
6966	* @param iBit The bit to test and set.
6967	*
6968	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
6969	* However, doing so will yield better performance as well as avoiding
6970	* traps accessing the last bits in the bitmap.
6971	*/
6972	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
6973	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
6974	#else
6975	DECLINLINE(bool) ASMBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
6976	{
6977	union { bool f; uint32_t u32; uint8_t u8; } rc;
6978	# if RT_INLINE_ASM_USES_INTRIN
6979	rc.u8 = _bittestandset((long RT_FAR *)pvBitmap, iBit);
6980
6981	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
6982	# if RT_INLINE_ASM_GNU_STYLE
6983	__asm__ __volatile__("btsl %2, %1\n\t"
6984	"setc %b0\n\t"
6985	"andl $1, %0\n\t"
6986	: "=q" (rc.u32)
6987	, "=m" ((volatile long RT_FAR )pvBitmap)
6988	: "Ir" (iBit)
6989	, "m" ((volatile long RT_FAR )pvBitmap)
6990	: "memory"
6991	, "cc");
6992	# else
6993	__asm
6994	{
6995	mov edx, [iBit]
6996	# ifdef RT_ARCH_AMD64
6997	mov rax, [pvBitmap]
6998	bts [rax], edx
6999	# else
7000	mov eax, [pvBitmap]
7001	bts [eax], edx
7002	# endif
7003	setc al
7004	and eax, 1
7005	mov [rc.u32], eax
7006	}
7007	# endif
7008
7009	# else
7010	int32_t offBitmap = iBit / 32;
7011	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
7012	rc.u32 = RT_LE2H_U32(ASMAtomicUoOrExU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31))))
7013	>> (iBit & 31);
7014	rc.u32 &= 1;
7015	# endif
7016	return rc.f;
7017	}
7018	#endif
7019
7020
7021	/**
7022	* Atomically tests and sets a bit in a bitmap, ordered.
7023	*
7024	* @returns true if the bit was set.
7025	* @returns false if the bit was clear.
7026	*
7027	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
7028	* aligned, otherwise the memory access isn't atomic!
7029	* @param iBit The bit to set.
7030	*
7031	* @remarks x86: Requires a 386 or later.
7032	*/
7033	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7034	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
7035	#else
7036	DECLINLINE(bool) ASMAtomicBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
7037	{
7038	union { bool f; uint32_t u32; uint8_t u8; } rc;
7039	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
7040	# if RT_INLINE_ASM_USES_INTRIN
7041	rc.u8 = _interlockedbittestandset((long RT_FAR *)pvBitmap, iBit);
7042	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7043	# if RT_INLINE_ASM_GNU_STYLE
7044	__asm__ __volatile__("lock; btsl %2, %1\n\t"
7045	"setc %b0\n\t"
7046	"andl $1, %0\n\t"
7047	: "=q" (rc.u32)
7048	, "=m" ((volatile long RT_FAR )pvBitmap)
7049	: "Ir" (iBit)
7050	, "m" ((volatile long RT_FAR )pvBitmap)
7051	: "memory"
7052	, "cc");
7053	# else
7054	__asm
7055	{
7056	mov edx, [iBit]
7057	# ifdef RT_ARCH_AMD64
7058	mov rax, [pvBitmap]
7059	lock bts [rax], edx
7060	# else
7061	mov eax, [pvBitmap]
7062	lock bts [eax], edx
7063	# endif
7064	setc al
7065	and eax, 1
7066	mov [rc.u32], eax
7067	}
7068	# endif
7069
7070	# else
7071	rc.u32 = RT_LE2H_U32(ASMAtomicOrExU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(RT_BIT_32(iBit & 31))))
7072	>> (iBit & 31);
7073	rc.u32 &= 1;
7074	# endif
7075	return rc.f;
7076	}
7077	#endif
7078
7079
7080	/**
7081	* Tests and clears a bit in a bitmap.
7082	*
7083	* @returns true if the bit was set.
7084	* @returns false if the bit was clear.
7085	*
7086	* @param pvBitmap Pointer to the bitmap (little endian).
7087	* @param iBit The bit to test and clear.
7088	*
7089	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
7090	* However, doing so will yield better performance as well as avoiding
7091	* traps accessing the last bits in the bitmap.
7092	*/
7093	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7094	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
7095	#else
7096	DECLINLINE(bool) ASMBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
7097	{
7098	union { bool f; uint32_t u32; uint8_t u8; } rc;
7099	# if RT_INLINE_ASM_USES_INTRIN
7100	rc.u8 = _bittestandreset((long RT_FAR *)pvBitmap, iBit);
7101
7102	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7103	# if RT_INLINE_ASM_GNU_STYLE
7104	__asm__ __volatile__("btrl %2, %1\n\t"
7105	"setc %b0\n\t"
7106	"andl $1, %0\n\t"
7107	: "=q" (rc.u32)
7108	, "=m" ((volatile long RT_FAR )pvBitmap)
7109	: "Ir" (iBit)
7110	, "m" ((volatile long RT_FAR )pvBitmap)
7111	: "memory"
7112	, "cc");
7113	# else
7114	__asm
7115	{
7116	mov edx, [iBit]
7117	# ifdef RT_ARCH_AMD64
7118	mov rax, [pvBitmap]
7119	btr [rax], edx
7120	# else
7121	mov eax, [pvBitmap]
7122	btr [eax], edx
7123	# endif
7124	setc al
7125	and eax, 1
7126	mov [rc.u32], eax
7127	}
7128	# endif
7129
7130	# else
7131	int32_t offBitmap = iBit / 32;
7132	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
7133	rc.u32 = RT_LE2H_U32(ASMAtomicUoAndExU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(~RT_BIT_32(iBit & 31))))
7134	>> (iBit & 31);
7135	rc.u32 &= 1;
7136	# endif
7137	return rc.f;
7138	}
7139	#endif
7140
7141
7142	/**
7143	* Atomically tests and clears a bit in a bitmap, ordered.
7144	*
7145	* @returns true if the bit was set.
7146	* @returns false if the bit was clear.
7147	*
7148	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
7149	* aligned, otherwise the memory access isn't atomic!
7150	* @param iBit The bit to test and clear.
7151	*
7152	* @remarks No memory barrier, take care on smp.
7153	* @remarks x86: Requires a 386 or later.
7154	*/
7155	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7156	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
7157	#else
7158	DECLINLINE(bool) ASMAtomicBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
7159	{
7160	union { bool f; uint32_t u32; uint8_t u8; } rc;
7161	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
7162	# if RT_INLINE_ASM_USES_INTRIN
7163	rc.u8 = _interlockedbittestandreset((long RT_FAR *)pvBitmap, iBit);
7164
7165	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7166	# if RT_INLINE_ASM_GNU_STYLE
7167	__asm__ __volatile__("lock; btrl %2, %1\n\t"
7168	"setc %b0\n\t"
7169	"andl $1, %0\n\t"
7170	: "=q" (rc.u32)
7171	, "=m" ((volatile long RT_FAR )pvBitmap)
7172	: "Ir" (iBit)
7173	, "m" ((volatile long RT_FAR )pvBitmap)
7174	: "memory"
7175	, "cc");
7176	# else
7177	__asm
7178	{
7179	mov edx, [iBit]
7180	# ifdef RT_ARCH_AMD64
7181	mov rax, [pvBitmap]
7182	lock btr [rax], edx
7183	# else
7184	mov eax, [pvBitmap]
7185	lock btr [eax], edx
7186	# endif
7187	setc al
7188	and eax, 1
7189	mov [rc.u32], eax
7190	}
7191	# endif
7192
7193	# else
7194	rc.u32 = RT_LE2H_U32(ASMAtomicAndExU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(~RT_BIT_32(iBit & 31))))
7195	>> (iBit & 31);
7196	rc.u32 &= 1;
7197	# endif
7198	return rc.f;
7199	}
7200	#endif
7201
7202
7203	/**
7204	* Tests and toggles a bit in a bitmap.
7205	*
7206	* @returns true if the bit was set.
7207	* @returns false if the bit was clear.
7208	*
7209	* @param pvBitmap Pointer to the bitmap (little endian).
7210	* @param iBit The bit to test and toggle.
7211	*
7212	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
7213	* However, doing so will yield better performance as well as avoiding
7214	* traps accessing the last bits in the bitmap.
7215	*/
7216	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7217	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
7218	#else
7219	DECLINLINE(bool) ASMBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
7220	{
7221	union { bool f; uint32_t u32; uint8_t u8; } rc;
7222	# if RT_INLINE_ASM_USES_INTRIN
7223	rc.u8 = _bittestandcomplement((long RT_FAR *)pvBitmap, iBit);
7224
7225	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7226	# if RT_INLINE_ASM_GNU_STYLE
7227	__asm__ __volatile__("btcl %2, %1\n\t"
7228	"setc %b0\n\t"
7229	"andl $1, %0\n\t"
7230	: "=q" (rc.u32)
7231	, "=m" ((volatile long RT_FAR )pvBitmap)
7232	: "Ir" (iBit)
7233	, "m" ((volatile long RT_FAR )pvBitmap)
7234	: "memory"
7235	, "cc");
7236	# else
7237	__asm
7238	{
7239	mov edx, [iBit]
7240	# ifdef RT_ARCH_AMD64
7241	mov rax, [pvBitmap]
7242	btc [rax], edx
7243	# else
7244	mov eax, [pvBitmap]
7245	btc [eax], edx
7246	# endif
7247	setc al
7248	and eax, 1
7249	mov [rc.u32], eax
7250	}
7251	# endif
7252
7253	# else
7254	int32_t offBitmap = iBit / 32;
7255	AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8);
7256	rc.u32 = RT_LE2H_U32(ASMAtomicUoXorExU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31))))
7257	>> (iBit & 31);
7258	rc.u32 &= 1;
7259	# endif
7260	return rc.f;
7261	}
7262	#endif
7263
7264
7265	/**
7266	* Atomically tests and toggles a bit in a bitmap, ordered.
7267	*
7268	* @returns true if the bit was set.
7269	* @returns false if the bit was clear.
7270	*
7271	* @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit
7272	* aligned, otherwise the memory access isn't atomic!
7273	* @param iBit The bit to test and toggle.
7274	*
7275	* @remarks x86: Requires a 386 or later.
7276	*/
7277	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM
7278	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
7279	#else
7280	DECLINLINE(bool) ASMAtomicBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
7281	{
7282	union { bool f; uint32_t u32; uint8_t u8; } rc;
7283	AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap));
7284	# if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7285	# if RT_INLINE_ASM_GNU_STYLE
7286	__asm__ __volatile__("lock; btcl %2, %1\n\t"
7287	"setc %b0\n\t"
7288	"andl $1, %0\n\t"
7289	: "=q" (rc.u32)
7290	, "=m" ((volatile long RT_FAR )pvBitmap)
7291	: "Ir" (iBit)
7292	, "m" ((volatile long RT_FAR )pvBitmap)
7293	: "memory"
7294	, "cc");
7295	# else
7296	__asm
7297	{
7298	mov edx, [iBit]
7299	# ifdef RT_ARCH_AMD64
7300	mov rax, [pvBitmap]
7301	lock btc [rax], edx
7302	# else
7303	mov eax, [pvBitmap]
7304	lock btc [eax], edx
7305	# endif
7306	setc al
7307	and eax, 1
7308	mov [rc.u32], eax
7309	}
7310	# endif
7311
7312	# else
7313	rc.u32 = RT_H2LE_U32(ASMAtomicXorExU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_LE2H_U32(RT_BIT_32(iBit & 31))))
7314	>> (iBit & 31);
7315	rc.u32 &= 1;
7316	# endif
7317	return rc.f;
7318	}
7319	#endif
7320
7321
7322	/**
7323	* Tests if a bit in a bitmap is set.
7324	*
7325	* @returns true if the bit is set.
7326	* @returns false if the bit is clear.
7327	*
7328	* @param pvBitmap Pointer to the bitmap (little endian).
7329	* @param iBit The bit to test.
7330	*
7331	* @remarks The 32-bit aligning of pvBitmap is not a strict requirement.
7332	* However, doing so will yield better performance as well as avoiding
7333	* traps accessing the last bits in the bitmap.
7334	*/
7335	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7336	RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTest(const volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO;
7337	#else
7338	DECLINLINE(bool) ASMBitTest(const volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF
7339	{
7340	union { bool f; uint32_t u32; uint8_t u8; } rc;
7341	# if RT_INLINE_ASM_USES_INTRIN
7342	rc.u32 = _bittest((long *)pvBitmap, iBit);
7343
7344	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7345	# if RT_INLINE_ASM_GNU_STYLE
7346
7347	__asm__ __volatile__("btl %2, %1\n\t"
7348	"setc %b0\n\t"
7349	"andl $1, %0\n\t"
7350	: "=q" (rc.u32)
7351	: "m" ((const volatile long RT_FAR )pvBitmap)
7352	, "Ir" (iBit)
7353	: "memory"
7354	, "cc");
7355	# else
7356	__asm
7357	{
7358	mov edx, [iBit]
7359	# ifdef RT_ARCH_AMD64
7360	mov rax, [pvBitmap]
7361	bt [rax], edx
7362	# else
7363	mov eax, [pvBitmap]
7364	bt [eax], edx
7365	# endif
7366	setc al
7367	and eax, 1
7368	mov [rc.u32], eax
7369	}
7370	# endif
7371
7372	# else
7373	int32_t offBitmap = iBit / 32;
7374	AssertRelease(!((uintptr_t)pvBitmap & (sizeof(uint32_t) - 1)));
7375	rc.u32 = RT_LE2H_U32(ASMAtomicUoReadU32(&((uint32_t volatile *)pvBitmap)[offBitmap])) >> (iBit & 31);
7376	rc.u32 &= 1;
7377	# endif
7378	return rc.f;
7379	}
7380	#endif
7381
7382
7383	#ifdef IPRT_INCLUDED_asm_mem_h
7384
7385	/**
7386	* Clears a bit range within a bitmap.
7387	*
7388	* @param pvBitmap Pointer to the bitmap (little endian).
7389	* @param iBitStart The First bit to clear.
7390	* @param iBitEnd The first bit not to clear.
7391	*/
7392	DECLINLINE(void) ASMBitClearRange(volatile void RT_FAR *pvBitmap, size_t iBitStart, size_t iBitEnd) RT_NOTHROW_DEF
7393	{
7394	if (iBitStart < iBitEnd)
7395	{
7396	uint32_t volatile RT_FAR pu32 = (volatile uint32_t RT_FAR )pvBitmap + (iBitStart >> 5);
7397	size_t iStart = iBitStart & ~(size_t)31;
7398	size_t iEnd = iBitEnd & ~(size_t)31;
7399	if (iStart == iEnd)
7400	*pu32 &= RT_H2LE_U32(((UINT32_C(1) << (iBitStart & 31)) - 1) \| ~((UINT32_C(1) << (iBitEnd & 31)) - 1));
7401	else
7402	{
7403	/* bits in first dword. */
7404	if (iBitStart & 31)
7405	{
7406	*pu32 &= RT_H2LE_U32((UINT32_C(1) << (iBitStart & 31)) - 1);
7407	pu32++;
7408	iBitStart = iStart + 32;
7409	}
7410
7411	/* whole dwords. */
7412	if (iBitStart != iEnd)
7413	ASMMemZero32(pu32, (iEnd - iBitStart) >> 3);
7414
7415	/* bits in last dword. */
7416	if (iBitEnd & 31)
7417	{
7418	pu32 = (volatile uint32_t RT_FAR *)pvBitmap + (iBitEnd >> 5);
7419	*pu32 &= RT_H2LE_U32(~((UINT32_C(1) << (iBitEnd & 31)) - 1));
7420	}
7421	}
7422	}
7423	}
7424
7425
7426	/**
7427	* Sets a bit range within a bitmap.
7428	*
7429	* @param pvBitmap Pointer to the bitmap (little endian).
7430	* @param iBitStart The First bit to set.
7431	* @param iBitEnd The first bit not to set.
7432	*/
7433	DECLINLINE(void) ASMBitSetRange(volatile void RT_FAR *pvBitmap, size_t iBitStart, size_t iBitEnd) RT_NOTHROW_DEF
7434	{
7435	if (iBitStart < iBitEnd)
7436	{
7437	uint32_t volatile RT_FAR pu32 = (volatile uint32_t RT_FAR )pvBitmap + (iBitStart >> 5);
7438	size_t iStart = iBitStart & ~(size_t)31;
7439	size_t iEnd = iBitEnd & ~(size_t)31;
7440	if (iStart == iEnd)
7441	*pu32 \|= RT_H2LE_U32(((UINT32_C(1) << (iBitEnd - iBitStart)) - 1) << (iBitStart & 31));
7442	else
7443	{
7444	/* bits in first dword. */
7445	if (iBitStart & 31)
7446	{
7447	*pu32 \|= RT_H2LE_U32(~((UINT32_C(1) << (iBitStart & 31)) - 1));
7448	pu32++;
7449	iBitStart = iStart + 32;
7450	}
7451
7452	/* whole dword. */
7453	if (iBitStart != iEnd)
7454	ASMMemFill32(pu32, (iEnd - iBitStart) >> 3, ~UINT32_C(0));
7455
7456	/* bits in last dword. */
7457	if (iBitEnd & 31)
7458	{
7459	pu32 = (volatile uint32_t RT_FAR *)pvBitmap + (iBitEnd >> 5);
7460	*pu32 \|= RT_H2LE_U32((UINT32_C(1) << (iBitEnd & 31)) - 1);
7461	}
7462	}
7463	}
7464	}
7465
7466	#endif /* IPRT_INCLUDED_asm_mem_h */
7467
7468	/**
7469	* Finds the first clear bit in a bitmap.
7470	*
7471	* @returns Index of the first zero bit.
7472	* @returns -1 if no clear bit was found.
7473	* @param pvBitmap Pointer to the bitmap (little endian).
7474	* @param cBits The number of bits in the bitmap. Multiple of 32.
7475	*/
7476	#if RT_INLINE_ASM_EXTERNAL \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
7477	DECLASM(int32_t) ASMBitFirstClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_PROTO;
7478	#else
7479	DECLINLINE(int32_t) ASMBitFirstClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_DEF
7480	{
7481	if (cBits)
7482	{
7483	int32_t iBit;
7484	# if RT_INLINE_ASM_GNU_STYLE
7485	RTCCUINTREG uEAX, uECX, uEDI;
7486	cBits = RT_ALIGN_32(cBits, 32);
7487	__asm__ __volatile__("repe; scasl\n\t"
7488	"je 1f\n\t"
7489	# ifdef RT_ARCH_AMD64
7490	"lea -4(%%rdi), %%rdi\n\t"
7491	"xorl (%%rdi), %%eax\n\t"
7492	"subq %5, %%rdi\n\t"
7493	# else
7494	"lea -4(%%edi), %%edi\n\t"
7495	"xorl (%%edi), %%eax\n\t"
7496	"subl %5, %%edi\n\t"
7497	# endif
7498	"shll $3, %%edi\n\t"
7499	"bsfl %%eax, %%edx\n\t"
7500	"addl %%edi, %%edx\n\t"
7501	"1:\t\n"
7502	: "=d" (iBit)
7503	, "=&c" (uECX)
7504	, "=&D" (uEDI)
7505	, "=&a" (uEAX)
7506	: "0" (0xffffffff)
7507	, "mr" (pvBitmap)
7508	, "1" (cBits >> 5)
7509	, "2" (pvBitmap)
7510	, "3" (0xffffffff)
7511	: "cc");
7512	# else
7513	cBits = RT_ALIGN_32(cBits, 32);
7514	__asm
7515	{
7516	# ifdef RT_ARCH_AMD64
7517	mov rdi, [pvBitmap]
7518	mov rbx, rdi
7519	# else
7520	mov edi, [pvBitmap]
7521	mov ebx, edi
7522	# endif
7523	mov edx, 0ffffffffh
7524	mov eax, edx
7525	mov ecx, [cBits]
7526	shr ecx, 5
7527	repe scasd
7528	je done
7529
7530	# ifdef RT_ARCH_AMD64
7531	lea rdi, [rdi - 4]
7532	xor eax, [rdi]
7533	sub rdi, rbx
7534	# else
7535	lea edi, [edi - 4]
7536	xor eax, [edi]
7537	sub edi, ebx
7538	# endif
7539	shl edi, 3
7540	bsf edx, eax
7541	add edx, edi
7542	done:
7543	mov [iBit], edx
7544	}
7545	# endif
7546	return iBit;
7547	}
7548	return -1;
7549	}
7550	#endif
7551
7552
7553	/**
7554	* Finds the next clear bit in a bitmap.
7555	*
7556	* @returns Index of the first zero bit.
7557	* @returns -1 if no clear bit was found.
7558	* @param pvBitmap Pointer to the bitmap (little endian).
7559	* @param cBits The number of bits in the bitmap. Multiple of 32.
7560	* @param iBitPrev The bit returned from the last search.
7561	* The search will start at iBitPrev + 1.
7562	*/
7563	#if RT_INLINE_ASM_EXTERNAL \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
7564	DECLASM(int) ASMBitNextClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_PROTO;
7565	#else
7566	DECLINLINE(int) ASMBitNextClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_DEF
7567	{
7568	const volatile uint32_t RT_FAR pau32Bitmap = (const volatile uint32_t RT_FAR )pvBitmap;
7569	int iBit = ++iBitPrev & 31;
7570	if (iBit)
7571	{
7572	/*
7573	* Inspect the 32-bit word containing the unaligned bit.
7574	*/
7575	uint32_t u32 = ~pau32Bitmap[iBitPrev / 32] >> iBit;
7576
7577	# if RT_INLINE_ASM_USES_INTRIN
7578	unsigned long ulBit = 0;
7579	if (_BitScanForward(&ulBit, u32))
7580	return ulBit + iBitPrev;
7581	# else
7582	# if RT_INLINE_ASM_GNU_STYLE
7583	__asm__ __volatile__("bsf %1, %0\n\t"
7584	"jnz 1f\n\t"
7585	"movl $-1, %0\n\t" /** @todo use conditional move for 64-bit? */
7586	"1:\n\t"
7587	: "=r" (iBit)
7588	: "r" (u32)
7589	: "cc");
7590	# else
7591	__asm
7592	{
7593	mov edx, [u32]
7594	bsf eax, edx
7595	jnz done
7596	mov eax, 0ffffffffh
7597	done:
7598	mov [iBit], eax
7599	}
7600	# endif
7601	if (iBit >= 0)
7602	return iBit + (int)iBitPrev;
7603	# endif
7604
7605	/*
7606	* Skip ahead and see if there is anything left to search.
7607	*/
7608	iBitPrev \|= 31;
7609	iBitPrev++;
7610	if (cBits <= (uint32_t)iBitPrev)
7611	return -1;
7612	}
7613
7614	/*
7615	* 32-bit aligned search, let ASMBitFirstClear do the dirty work.
7616	*/
7617	iBit = ASMBitFirstClear(&pau32Bitmap[iBitPrev / 32], cBits - iBitPrev);
7618	if (iBit >= 0)
7619	iBit += iBitPrev;
7620	return iBit;
7621	}
7622	#endif
7623
7624
7625	/**
7626	* Finds the first set bit in a bitmap.
7627	*
7628	* @returns Index of the first set bit.
7629	* @returns -1 if no clear bit was found.
7630	* @param pvBitmap Pointer to the bitmap (little endian).
7631	* @param cBits The number of bits in the bitmap. Multiple of 32.
7632	*/
7633	#if RT_INLINE_ASM_EXTERNAL \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
7634	DECLASM(int32_t) ASMBitFirstSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_PROTO;
7635	#else
7636	DECLINLINE(int32_t) ASMBitFirstSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_DEF
7637	{
7638	if (cBits)
7639	{
7640	int32_t iBit;
7641	# if RT_INLINE_ASM_GNU_STYLE
7642	RTCCUINTREG uEAX, uECX, uEDI;
7643	cBits = RT_ALIGN_32(cBits, 32);
7644	__asm__ __volatile__("repe; scasl\n\t"
7645	"je 1f\n\t"
7646	# ifdef RT_ARCH_AMD64
7647	"lea -4(%%rdi), %%rdi\n\t"
7648	"movl (%%rdi), %%eax\n\t"
7649	"subq %5, %%rdi\n\t"
7650	# else
7651	"lea -4(%%edi), %%edi\n\t"
7652	"movl (%%edi), %%eax\n\t"
7653	"subl %5, %%edi\n\t"
7654	# endif
7655	"shll $3, %%edi\n\t"
7656	"bsfl %%eax, %%edx\n\t"
7657	"addl %%edi, %%edx\n\t"
7658	"1:\t\n"
7659	: "=d" (iBit)
7660	, "=&c" (uECX)
7661	, "=&D" (uEDI)
7662	, "=&a" (uEAX)
7663	: "0" (0xffffffff)
7664	, "mr" (pvBitmap)
7665	, "1" (cBits >> 5)
7666	, "2" (pvBitmap)
7667	, "3" (0)
7668	: "cc");
7669	# else
7670	cBits = RT_ALIGN_32(cBits, 32);
7671	__asm
7672	{
7673	# ifdef RT_ARCH_AMD64
7674	mov rdi, [pvBitmap]
7675	mov rbx, rdi
7676	# else
7677	mov edi, [pvBitmap]
7678	mov ebx, edi
7679	# endif
7680	mov edx, 0ffffffffh
7681	xor eax, eax
7682	mov ecx, [cBits]
7683	shr ecx, 5
7684	repe scasd
7685	je done
7686	# ifdef RT_ARCH_AMD64
7687	lea rdi, [rdi - 4]
7688	mov eax, [rdi]
7689	sub rdi, rbx
7690	# else
7691	lea edi, [edi - 4]
7692	mov eax, [edi]
7693	sub edi, ebx
7694	# endif
7695	shl edi, 3
7696	bsf edx, eax
7697	add edx, edi
7698	done:
7699	mov [iBit], edx
7700	}
7701	# endif
7702	return iBit;
7703	}
7704	return -1;
7705	}
7706	#endif
7707
7708
7709	/**
7710	* Finds the next set bit in a bitmap.
7711	*
7712	* @returns Index of the next set bit.
7713	* @returns -1 if no set bit was found.
7714	* @param pvBitmap Pointer to the bitmap (little endian).
7715	* @param cBits The number of bits in the bitmap. Multiple of 32.
7716	* @param iBitPrev The bit returned from the last search.
7717	* The search will start at iBitPrev + 1.
7718	*/
7719	#if RT_INLINE_ASM_EXTERNAL \|\| (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
7720	DECLASM(int) ASMBitNextSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_PROTO;
7721	#else
7722	DECLINLINE(int) ASMBitNextSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_DEF
7723	{
7724	const volatile uint32_t RT_FAR pau32Bitmap = (const volatile uint32_t RT_FAR )pvBitmap;
7725	int iBit = ++iBitPrev & 31;
7726	if (iBit)
7727	{
7728	/*
7729	* Inspect the 32-bit word containing the unaligned bit.
7730	*/
7731	uint32_t u32 = pau32Bitmap[iBitPrev / 32] >> iBit;
7732
7733	# if RT_INLINE_ASM_USES_INTRIN
7734	unsigned long ulBit = 0;
7735	if (_BitScanForward(&ulBit, u32))
7736	return ulBit + iBitPrev;
7737	# else
7738	# if RT_INLINE_ASM_GNU_STYLE
7739	__asm__ __volatile__("bsf %1, %0\n\t"
7740	"jnz 1f\n\t" /** @todo use conditional move for 64-bit? */
7741	"movl $-1, %0\n\t"
7742	"1:\n\t"
7743	: "=r" (iBit)
7744	: "r" (u32)
7745	: "cc");
7746	# else
7747	__asm
7748	{
7749	mov edx, [u32]
7750	bsf eax, edx
7751	jnz done
7752	mov eax, 0ffffffffh
7753	done:
7754	mov [iBit], eax
7755	}
7756	# endif
7757	if (iBit >= 0)
7758	return iBit + (int)iBitPrev;
7759	# endif
7760
7761	/*
7762	* Skip ahead and see if there is anything left to search.
7763	*/
7764	iBitPrev \|= 31;
7765	iBitPrev++;
7766	if (cBits <= (uint32_t)iBitPrev)
7767	return -1;
7768	}
7769
7770	/*
7771	* 32-bit aligned search, let ASMBitFirstClear do the dirty work.
7772	*/
7773	iBit = ASMBitFirstSet(&pau32Bitmap[iBitPrev / 32], cBits - iBitPrev);
7774	if (iBit >= 0)
7775	iBit += iBitPrev;
7776	return iBit;
7777	}
7778	#endif
7779
7780	/** @} */
7781
7782
7783	/** @defgroup grp_inline_bits Bitmap Operations
7784	* @{
7785	*/
7786
7787	/**
7788	* Finds the first bit which is set in the given 32-bit integer.
7789	* Bits are numbered from 1 (least significant) to 32.
7790	*
7791	* @returns index [1..32] of the first set bit.
7792	* @returns 0 if all bits are cleared.
7793	* @param u32 Integer to search for set bits.
7794	* @remarks Similar to ffs() in BSD.
7795	*/
7796	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7797	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitFirstSetU32(uint32_t u32) RT_NOTHROW_PROTO;
7798	#else
7799	DECLINLINE(unsigned) ASMBitFirstSetU32(uint32_t u32) RT_NOTHROW_DEF
7800	{
7801	# if RT_INLINE_ASM_USES_INTRIN
7802	unsigned long iBit;
7803	if (_BitScanForward(&iBit, u32))
7804	iBit++;
7805	else
7806	iBit = 0;
7807
7808	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7809	# if RT_INLINE_ASM_GNU_STYLE
7810	uint32_t iBit;
7811	__asm__ __volatile__("bsf %1, %0\n\t"
7812	"jnz 1f\n\t"
7813	"xorl %0, %0\n\t"
7814	"jmp 2f\n"
7815	"1:\n\t"
7816	"incl %0\n"
7817	"2:\n\t"
7818	: "=r" (iBit)
7819	: "rm" (u32)
7820	: "cc");
7821	# else
7822	uint32_t iBit;
7823	_asm
7824	{
7825	bsf eax, [u32]
7826	jnz found
7827	xor eax, eax
7828	jmp done
7829	found:
7830	inc eax
7831	done:
7832	mov [iBit], eax
7833	}
7834	# endif
7835
7836	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
7837	/*
7838	* Using the "count leading zeros (clz)" instruction here because there
7839	* is no dedicated instruction to get the first set bit.
7840	* Need to reverse the bits in the value with "rbit" first because
7841	* "clz" starts counting from the most significant bit.
7842	*/
7843	uint32_t iBit;
7844	__asm__ __volatile__(
7845	# if defined(RT_ARCH_ARM64)
7846	"rbit %w[uVal], %w[uVal]\n\t"
7847	"clz %w[iBit], %w[uVal]\n\t"
7848	# else
7849	"rbit %[uVal], %[uVal]\n\t"
7850	"clz %[iBit], %[uVal]\n\t"
7851	# endif
7852	: [uVal] "=r" (u32)
7853	, [iBit] "=r" (iBit)
7854	: "[uVal]" (u32));
7855	if (iBit != 32)
7856	iBit++;
7857	else
7858	iBit = 0; /* No bit set. */
7859
7860	# else
7861	# error "Port me"
7862	# endif
7863	return iBit;
7864	}
7865	#endif
7866
7867
7868	/**
7869	* Finds the first bit which is set in the given 32-bit integer.
7870	* Bits are numbered from 1 (least significant) to 32.
7871	*
7872	* @returns index [1..32] of the first set bit.
7873	* @returns 0 if all bits are cleared.
7874	* @param i32 Integer to search for set bits.
7875	* @remark Similar to ffs() in BSD.
7876	*/
7877	DECLINLINE(unsigned) ASMBitFirstSetS32(int32_t i32) RT_NOTHROW_DEF
7878	{
7879	return ASMBitFirstSetU32((uint32_t)i32);
7880	}
7881
7882
7883	/**
7884	* Finds the first bit which is set in the given 64-bit integer.
7885	*
7886	* Bits are numbered from 1 (least significant) to 64.
7887	*
7888	* @returns index [1..64] of the first set bit.
7889	* @returns 0 if all bits are cleared.
7890	* @param u64 Integer to search for set bits.
7891	* @remarks Similar to ffs() in BSD.
7892	*/
7893	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7894	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitFirstSetU64(uint64_t u64) RT_NOTHROW_PROTO;
7895	#else
7896	DECLINLINE(unsigned) ASMBitFirstSetU64(uint64_t u64) RT_NOTHROW_DEF
7897	{
7898	# if RT_INLINE_ASM_USES_INTRIN
7899	unsigned long iBit;
7900	# if ARCH_BITS == 64
7901	if (_BitScanForward64(&iBit, u64))
7902	iBit++;
7903	else
7904	iBit = 0;
7905	# else
7906	if (_BitScanForward(&iBit, (uint32_t)u64))
7907	iBit++;
7908	else if (_BitScanForward(&iBit, (uint32_t)(u64 >> 32)))
7909	iBit += 33;
7910	else
7911	iBit = 0;
7912	# endif
7913
7914	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
7915	uint64_t iBit;
7916	__asm__ __volatile__("bsfq %1, %0\n\t"
7917	"jnz 1f\n\t"
7918	"xorl %k0, %k0\n\t"
7919	"jmp 2f\n"
7920	"1:\n\t"
7921	"incl %k0\n"
7922	"2:\n\t"
7923	: "=r" (iBit)
7924	: "rm" (u64)
7925	: "cc");
7926
7927	# elif defined(RT_ARCH_ARM64)
7928	uint64_t iBit;
7929	__asm__ __volatile__("rbit %[uVal], %[uVal]\n\t"
7930	"clz %[iBit], %[uVal]\n\t"
7931	: [uVal] "=r" (u64)
7932	, [iBit] "=r" (iBit)
7933	: "[uVal]" (u64));
7934	if (iBit != 64)
7935	iBit++;
7936	else
7937	iBit = 0; /* No bit set. */
7938
7939	# else
7940	unsigned iBit = ASMBitFirstSetU32((uint32_t)u64);
7941	if (!iBit)
7942	{
7943	iBit = ASMBitFirstSetU32((uint32_t)(u64 >> 32));
7944	if (iBit)
7945	iBit += 32;
7946	}
7947	# endif
7948	return (unsigned)iBit;
7949	}
7950	#endif
7951
7952
7953	/**
7954	* Finds the first bit which is set in the given 16-bit integer.
7955	*
7956	* Bits are numbered from 1 (least significant) to 16.
7957	*
7958	* @returns index [1..16] of the first set bit.
7959	* @returns 0 if all bits are cleared.
7960	* @param u16 Integer to search for set bits.
7961	* @remarks For 16-bit bs3kit code.
7962	*/
7963	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7964	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitFirstSetU16(uint16_t u16) RT_NOTHROW_PROTO;
7965	#else
7966	DECLINLINE(unsigned) ASMBitFirstSetU16(uint16_t u16) RT_NOTHROW_DEF
7967	{
7968	return ASMBitFirstSetU32((uint32_t)u16);
7969	}
7970	#endif
7971
7972
7973	/**
7974	* Finds the last bit which is set in the given 32-bit integer.
7975	* Bits are numbered from 1 (least significant) to 32.
7976	*
7977	* @returns index [1..32] of the last set bit.
7978	* @returns 0 if all bits are cleared.
7979	* @param u32 Integer to search for set bits.
7980	* @remark Similar to fls() in BSD.
7981	*/
7982	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
7983	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitLastSetU32(uint32_t u32) RT_NOTHROW_PROTO;
7984	#else
7985	DECLINLINE(unsigned) ASMBitLastSetU32(uint32_t u32) RT_NOTHROW_DEF
7986	{
7987	# if RT_INLINE_ASM_USES_INTRIN
7988	unsigned long iBit;
7989	if (_BitScanReverse(&iBit, u32))
7990	iBit++;
7991	else
7992	iBit = 0;
7993
7994	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
7995	# if RT_INLINE_ASM_GNU_STYLE
7996	uint32_t iBit;
7997	__asm__ __volatile__("bsrl %1, %0\n\t"
7998	"jnz 1f\n\t"
7999	"xorl %0, %0\n\t"
8000	"jmp 2f\n"
8001	"1:\n\t"
8002	"incl %0\n"
8003	"2:\n\t"
8004	: "=r" (iBit)
8005	: "rm" (u32)
8006	: "cc");
8007	# else
8008	uint32_t iBit;
8009	_asm
8010	{
8011	bsr eax, [u32]
8012	jnz found
8013	xor eax, eax
8014	jmp done
8015	found:
8016	inc eax
8017	done:
8018	mov [iBit], eax
8019	}
8020	# endif
8021
8022	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
8023	uint32_t iBit;
8024	__asm__ __volatile__(
8025	# if defined(RT_ARCH_ARM64)
8026	"clz %w[iBit], %w[uVal]\n\t"
8027	# else
8028	"clz %[iBit], %[uVal]\n\t"
8029	# endif
8030	: [iBit] "=r" (iBit)
8031	: [uVal] "r" (u32));
8032	iBit = 32 - iBit;
8033
8034	# else
8035	# error "Port me"
8036	# endif
8037	return iBit;
8038	}
8039	#endif
8040
8041
8042	/**
8043	* Finds the last bit which is set in the given 32-bit integer.
8044	* Bits are numbered from 1 (least significant) to 32.
8045	*
8046	* @returns index [1..32] of the last set bit.
8047	* @returns 0 if all bits are cleared.
8048	* @param i32 Integer to search for set bits.
8049	* @remark Similar to fls() in BSD.
8050	*/
8051	DECLINLINE(unsigned) ASMBitLastSetS32(int32_t i32) RT_NOTHROW_DEF
8052	{
8053	return ASMBitLastSetU32((uint32_t)i32);
8054	}
8055
8056
8057	/**
8058	* Finds the last bit which is set in the given 64-bit integer.
8059	*
8060	* Bits are numbered from 1 (least significant) to 64.
8061	*
8062	* @returns index [1..64] of the last set bit.
8063	* @returns 0 if all bits are cleared.
8064	* @param u64 Integer to search for set bits.
8065	* @remark Similar to fls() in BSD.
8066	*/
8067	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
8068	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitLastSetU64(uint64_t u64) RT_NOTHROW_PROTO;
8069	#else
8070	DECLINLINE(unsigned) ASMBitLastSetU64(uint64_t u64) RT_NOTHROW_DEF
8071	{
8072	# if RT_INLINE_ASM_USES_INTRIN
8073	unsigned long iBit;
8074	# if ARCH_BITS == 64
8075	if (_BitScanReverse64(&iBit, u64))
8076	iBit++;
8077	else
8078	iBit = 0;
8079	# else
8080	if (_BitScanReverse(&iBit, (uint32_t)(u64 >> 32)))
8081	iBit += 33;
8082	else if (_BitScanReverse(&iBit, (uint32_t)u64))
8083	iBit++;
8084	else
8085	iBit = 0;
8086	# endif
8087
8088	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
8089	uint64_t iBit;
8090	__asm__ __volatile__("bsrq %1, %0\n\t"
8091	"jnz 1f\n\t"
8092	"xorl %k0, %k0\n\t"
8093	"jmp 2f\n"
8094	"1:\n\t"
8095	"incl %k0\n"
8096	"2:\n\t"
8097	: "=r" (iBit)
8098	: "rm" (u64)
8099	: "cc");
8100
8101	# elif defined(RT_ARCH_ARM64)
8102	uint64_t iBit;
8103	__asm__ __volatile__("clz %[iBit], %[uVal]\n\t"
8104	: [iBit] "=r" (iBit)
8105	: [uVal] "r" (u64));
8106	iBit = 64 - iBit;
8107
8108	# else
8109	unsigned iBit = ASMBitLastSetU32((uint32_t)(u64 >> 32));
8110	if (iBit)
8111	iBit += 32;
8112	else
8113	iBit = ASMBitLastSetU32((uint32_t)u64);
8114	# endif
8115	return (unsigned)iBit;
8116	}
8117	#endif
8118
8119
8120	/**
8121	* Finds the last bit which is set in the given 16-bit integer.
8122	*
8123	* Bits are numbered from 1 (least significant) to 16.
8124	*
8125	* @returns index [1..16] of the last set bit.
8126	* @returns 0 if all bits are cleared.
8127	* @param u16 Integer to search for set bits.
8128	* @remarks For 16-bit bs3kit code.
8129	*/
8130	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
8131	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitLastSetU16(uint16_t u16) RT_NOTHROW_PROTO;
8132	#else
8133	DECLINLINE(unsigned) ASMBitLastSetU16(uint16_t u16) RT_NOTHROW_DEF
8134	{
8135	return ASMBitLastSetU32((uint32_t)u16);
8136	}
8137	#endif
8138
8139
8140	/**
8141	* Count the number of leading zero bits in the given 32-bit integer.
8142	*
8143	* The counting starts with the most significate bit.
8144	*
8145	* @returns Number of most significant zero bits.
8146	* @returns 32 if all bits are cleared.
8147	* @param u32 Integer to consider.
8148	* @remarks Similar to __builtin_clz() in gcc, except defined zero input result.
8149	*/
8150	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
8151	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountLeadingZerosU32(uint32_t u32) RT_NOTHROW_PROTO;
8152	#else
8153	DECLINLINE(unsigned) ASMCountLeadingZerosU32(uint32_t u32) RT_NOTHROW_DEF
8154	{
8155	# if RT_INLINE_ASM_USES_INTRIN
8156	unsigned long iBit;
8157	if (!_BitScanReverse(&iBit, u32))
8158	return 32;
8159	return 31 - (unsigned)iBit;
8160
8161	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
8162	uint32_t iBit;
8163	# if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) && 0 /* significantly slower on 10980xe; 929 vs 237 ps/call */
8164	__asm__ __volatile__("bsrl %1, %0\n\t"
8165	"cmovzl %2, %0\n\t"
8166	: "=&r" (iBit)
8167	: "rm" (u32)
8168	, "rm" ((int32_t)-1)
8169	: "cc");
8170	# elif RT_INLINE_ASM_GNU_STYLE
8171	__asm__ __volatile__("bsr %1, %0\n\t"
8172	"jnz 1f\n\t"
8173	"mov $-1, %0\n\t"
8174	"1:\n\t"
8175	: "=r" (iBit)
8176	: "rm" (u32)
8177	: "cc");
8178	# else
8179	_asm
8180	{
8181	bsr eax, [u32]
8182	jnz found
8183	mov eax, -1
8184	found:
8185	mov [iBit], eax
8186	}
8187	# endif
8188	return 31 - iBit;
8189
8190	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
8191	uint32_t iBit;
8192	__asm__ __volatile__(
8193	# if defined(RT_ARCH_ARM64)
8194	"clz %w[iBit], %w[uVal]\n\t"
8195	# else
8196	"clz %[iBit], %[uVal]\n\t"
8197	# endif
8198	: [uVal] "=r" (u32)
8199	, [iBit] "=r" (iBit)
8200	: "[uVal]" (u32));
8201	return iBit;
8202
8203	# elif defined(__GNUC__)
8204	AssertCompile(sizeof(u32) == sizeof(unsigned int));
8205	return u32 ? __builtin_clz(u32) : 32;
8206
8207	# else
8208	# error "Port me"
8209	# endif
8210	}
8211	#endif
8212
8213
8214	/**
8215	* Count the number of leading zero bits in the given 64-bit integer.
8216	*
8217	* The counting starts with the most significate bit.
8218	*
8219	* @returns Number of most significant zero bits.
8220	* @returns 64 if all bits are cleared.
8221	* @param u64 Integer to consider.
8222	* @remarks Similar to __builtin_clzl() in gcc, except defined zero input
8223	* result.
8224	*/
8225	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
8226	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountLeadingZerosU64(uint64_t u64) RT_NOTHROW_PROTO;
8227	#else
8228	DECLINLINE(unsigned) ASMCountLeadingZerosU64(uint64_t u64) RT_NOTHROW_DEF
8229	{
8230	# if RT_INLINE_ASM_USES_INTRIN
8231	unsigned long iBit;
8232	# if ARCH_BITS == 64
8233	if (_BitScanReverse64(&iBit, u64))
8234	return 63 - (unsigned)iBit;
8235	# else
8236	if (_BitScanReverse(&iBit, (uint32_t)(u64 >> 32)))
8237	return 31 - (unsigned)iBit;
8238	if (_BitScanReverse(&iBit, (uint32_t)u64))
8239	return 63 - (unsigned)iBit;
8240	# endif
8241	return 64;
8242
8243	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
8244	uint64_t iBit;
8245	# if 0 /* 10980xe benchmark: 932 ps/call - the slower variant */
8246	__asm__ __volatile__("bsrq %1, %0\n\t"
8247	"cmovzq %2, %0\n\t"
8248	: "=&r" (iBit)
8249	: "rm" (u64)
8250	, "rm" ((int64_t)-1)
8251	: "cc");
8252	# else /* 10980xe benchmark: 262 ps/call */
8253	__asm__ __volatile__("bsrq %1, %0\n\t"
8254	"jnz 1f\n\t"
8255	"mov $-1, %0\n\t"
8256	"1:\n\t"
8257	: "=&r" (iBit)
8258	: "rm" (u64)
8259	: "cc");
8260	# endif
8261	return 63 - (unsigned)iBit;
8262
8263	# elif defined(RT_ARCH_ARM64)
8264	uint64_t iBit;
8265	__asm__ __volatile__("clz %[iBit], %[uVal]\n\t"
8266	: [uVal] "=r" (u64)
8267	, [iBit] "=r" (iBit)
8268	: "[uVal]" (u64));
8269	return (unsigned)iBit;
8270
8271	# elif defined(__GNUC__) && ARCH_BITS == 64
8272	AssertCompile(sizeof(u64) == sizeof(unsigned long));
8273	return u64 ? __builtin_clzl(u64) : 64;
8274
8275	# else
8276	unsigned iBit = ASMCountLeadingZerosU32((uint32_t)(u64 >> 32));
8277	if (iBit == 32)
8278	iBit = ASMCountLeadingZerosU32((uint32_t)u64) + 32;
8279	return iBit;
8280	# endif
8281	}
8282	#endif
8283
8284
8285	/**
8286	* Count the number of leading zero bits in the given 16-bit integer.
8287	*
8288	* The counting starts with the most significate bit.
8289	*
8290	* @returns Number of most significant zero bits.
8291	* @returns 16 if all bits are cleared.
8292	* @param u16 Integer to consider.
8293	*/
8294	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
8295	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountLeadingZerosU16(uint16_t u16) RT_NOTHROW_PROTO;
8296	#else
8297	DECLINLINE(unsigned) ASMCountLeadingZerosU16(uint16_t u16) RT_NOTHROW_DEF
8298	{
8299	# if RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)) && 0 /* slower (10980xe: 987 vs 292 ps/call) */
8300	uint16_t iBit;
8301	__asm__ __volatile__("bsrw %1, %0\n\t"
8302	"jnz 1f\n\t"
8303	"mov $-1, %0\n\t"
8304	"1:\n\t"
8305	: "=r" (iBit)
8306	: "rm" (u16)
8307	: "cc");
8308	return 15 - (int16_t)iBit;
8309	# else
8310	return ASMCountLeadingZerosU32((uint32_t)u16) - 16;
8311	# endif
8312	}
8313	#endif
8314
8315
8316	/**
8317	* Count the number of trailing zero bits in the given 32-bit integer.
8318	*
8319	* The counting starts with the least significate bit, i.e. the zero bit.
8320	*
8321	* @returns Number of least significant zero bits.
8322	* @returns 32 if all bits are cleared.
8323	* @param u32 Integer to consider.
8324	* @remarks Similar to __builtin_ctz() in gcc, except defined zero input result.
8325	*/
8326	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
8327	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountTrailingZerosU32(uint32_t u32) RT_NOTHROW_PROTO;
8328	#else
8329	DECLINLINE(unsigned) ASMCountTrailingZerosU32(uint32_t u32) RT_NOTHROW_DEF
8330	{
8331	# if RT_INLINE_ASM_USES_INTRIN
8332	unsigned long iBit;
8333	if (!_BitScanForward(&iBit, u32))
8334	return 32;
8335	return (unsigned)iBit;
8336
8337	# elif defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
8338	uint32_t iBit;
8339	# if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) && 0 /* significantly slower on 10980xe; 932 vs 240 ps/call */
8340	__asm__ __volatile__("bsfl %1, %0\n\t"
8341	"cmovzl %2, %0\n\t"
8342	: "=&r" (iBit)
8343	: "rm" (u32)
8344	, "rm" ((int32_t)32)
8345	: "cc");
8346	# elif RT_INLINE_ASM_GNU_STYLE
8347	__asm__ __volatile__("bsfl %1, %0\n\t"
8348	"jnz 1f\n\t"
8349	"mov $32, %0\n\t"
8350	"1:\n\t"
8351	: "=r" (iBit)
8352	: "rm" (u32)
8353	: "cc");
8354	# else
8355	_asm
8356	{
8357	bsf eax, [u32]
8358	jnz found
8359	mov eax, 32
8360	found:
8361	mov [iBit], eax
8362	}
8363	# endif
8364	return iBit;
8365
8366	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
8367	/* Invert the bits and use clz. */
8368	uint32_t iBit;
8369	__asm__ __volatile__(
8370	# if defined(RT_ARCH_ARM64)
8371	"rbit %w[uVal], %w[uVal]\n\t"
8372	"clz %w[iBit], %w[uVal]\n\t"
8373	# else
8374	"rbit %[uVal], %[uVal]\n\t"
8375	"clz %[iBit], %[uVal]\n\t"
8376	# endif
8377	: [uVal] "=r" (u32)
8378	, [iBit] "=r" (iBit)
8379	: "[uVal]" (u32));
8380	return iBit;
8381
8382	# elif defined(__GNUC__)
8383	AssertCompile(sizeof(u32) == sizeof(unsigned int));
8384	return u32 ? __builtin_ctz(u32) : 32;
8385
8386	# else
8387	# error "Port me"
8388	# endif
8389	}
8390	#endif
8391
8392
8393	/**
8394	* Count the number of trailing zero bits in the given 64-bit integer.
8395	*
8396	* The counting starts with the least significate bit.
8397	*
8398	* @returns Number of least significant zero bits.
8399	* @returns 64 if all bits are cleared.
8400	* @param u64 Integer to consider.
8401	* @remarks Similar to __builtin_ctzl() in gcc, except defined zero input
8402	* result.
8403	*/
8404	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
8405	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountTrailingZerosU64(uint64_t u64) RT_NOTHROW_PROTO;
8406	#else
8407	DECLINLINE(unsigned) ASMCountTrailingZerosU64(uint64_t u64) RT_NOTHROW_DEF
8408	{
8409	# if RT_INLINE_ASM_USES_INTRIN
8410	unsigned long iBit;
8411	# if ARCH_BITS == 64
8412	if (_BitScanForward64(&iBit, u64))
8413	return (unsigned)iBit;
8414	# else
8415	if (_BitScanForward(&iBit, (uint32_t)u64))
8416	return (unsigned)iBit;
8417	if (_BitScanForward(&iBit, (uint32_t)(u64 >> 32)))
8418	return (unsigned)iBit + 32;
8419	# endif
8420	return 64;
8421
8422	# elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
8423	uint64_t iBit;
8424	# if 0 /* 10980xe benchmark: 932 ps/call - the slower variant */
8425	__asm__ __volatile__("bsfq %1, %0\n\t"
8426	"cmovzq %2, %0\n\t"
8427	: "=&r" (iBit)
8428	: "rm" (u64)
8429	, "rm" ((int64_t)64)
8430	: "cc");
8431	# else /* 10980xe benchmark: 262 ps/call */
8432	__asm__ __volatile__("bsfq %1, %0\n\t"
8433	"jnz 1f\n\t"
8434	"mov $64, %0\n\t"
8435	"1:\n\t"
8436	: "=&r" (iBit)
8437	: "rm" (u64)
8438	: "cc");
8439	# endif
8440	return (unsigned)iBit;
8441
8442	# elif defined(RT_ARCH_ARM64)
8443	/* Invert the bits and use clz. */
8444	uint64_t iBit;
8445	__asm__ __volatile__("rbit %[uVal], %[uVal]\n\t"
8446	"clz %[iBit], %[uVal]\n\t"
8447	: [uVal] "=r" (u64)
8448	, [iBit] "=r" (iBit)
8449	: "[uVal]" (u64));
8450	return (unsigned)iBit;
8451
8452	# elif defined(__GNUC__) && ARCH_BITS == 64
8453	AssertCompile(sizeof(u64) == sizeof(unsigned long));
8454	return u64 ? __builtin_ctzl(u64) : 64;
8455
8456	# else
8457	unsigned iBit = ASMCountTrailingZerosU32((uint32_t)u64);
8458	if (iBit == 32)
8459	iBit = ASMCountTrailingZerosU32((uint32_t)(u64 >> 32)) + 32;
8460	return iBit;
8461	# endif
8462	}
8463	#endif
8464
8465
8466	/**
8467	* Count the number of trailing zero bits in the given 16-bit integer.
8468	*
8469	* The counting starts with the most significate bit.
8470	*
8471	* @returns Number of most significant zero bits.
8472	* @returns 16 if all bits are cleared.
8473	* @param u16 Integer to consider.
8474	*/
8475	#if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN
8476	RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountTrailingZerosU16(uint16_t u16) RT_NOTHROW_PROTO;
8477	#else
8478	DECLINLINE(unsigned) ASMCountTrailingZerosU16(uint16_t u16) RT_NOTHROW_DEF
8479	{
8480	# if RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)) && 0 /* slower (10980xe: 992 vs 349 ps/call) */
8481	uint16_t iBit;
8482	__asm__ __volatile__("bsfw %1, %0\n\t"
8483	"jnz 1f\n\t"
8484	"mov $16, %0\n\t"
8485	"1:\n\t"
8486	: "=r" (iBit)
8487	: "rm" (u16)
8488	: "cc");
8489	return iBit;
8490	# else
8491	return ASMCountTrailingZerosU32((uint32_t)u16 \| UINT32_C(0x10000));
8492	#endif
8493	}
8494	#endif
8495
8496
8497	/**
8498	* Rotate 32-bit unsigned value to the left by @a cShift.
8499	*
8500	* @returns Rotated value.
8501	* @param u32 The value to rotate.
8502	* @param cShift How many bits to rotate by.
8503	*/
8504	#ifdef __WATCOMC__
8505	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMRotateLeftU32(uint32_t u32, unsigned cShift) RT_NOTHROW_PROTO;
8506	#else
8507	DECLINLINE(uint32_t) ASMRotateLeftU32(uint32_t u32, uint32_t cShift) RT_NOTHROW_DEF
8508	{
8509	# if RT_INLINE_ASM_USES_INTRIN
8510	return _rotl(u32, cShift);
8511
8512	# elif RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
8513	__asm__ __volatile__("roll %b1, %0" : "=g" (u32) : "Ic" (cShift), "0" (u32) : "cc");
8514	return u32;
8515
8516	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
8517	__asm__ __volatile__(
8518	# if defined(RT_ARCH_ARM64)
8519	"ror %w[uRet], %w[uVal], %w[cShift]\n\t"
8520	# else
8521	"ror %[uRet], %[uVal], %[cShift]\n\t"
8522	# endif
8523	: [uRet] "=r" (u32)
8524	: [uVal] "[uRet]" (u32)
8525	, [cShift] "r" (32 - (cShift & 31))); /** @todo there is an immediate form here */
8526	return u32;
8527
8528	# else
8529	cShift &= 31;
8530	return (u32 << cShift) \| (u32 >> (32 - cShift));
8531	# endif
8532	}
8533	#endif
8534
8535
8536	/**
8537	* Rotate 32-bit unsigned value to the right by @a cShift.
8538	*
8539	* @returns Rotated value.
8540	* @param u32 The value to rotate.
8541	* @param cShift How many bits to rotate by.
8542	*/
8543	#ifdef __WATCOMC__
8544	RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMRotateRightU32(uint32_t u32, unsigned cShift) RT_NOTHROW_PROTO;
8545	#else
8546	DECLINLINE(uint32_t) ASMRotateRightU32(uint32_t u32, uint32_t cShift) RT_NOTHROW_DEF
8547	{
8548	# if RT_INLINE_ASM_USES_INTRIN
8549	return _rotr(u32, cShift);
8550
8551	# elif RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
8552	__asm__ __volatile__("rorl %b1, %0" : "=g" (u32) : "Ic" (cShift), "0" (u32) : "cc");
8553	return u32;
8554
8555	# elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
8556	__asm__ __volatile__(
8557	# if defined(RT_ARCH_ARM64)
8558	"ror %w[uRet], %w[uVal], %w[cShift]\n\t"
8559	# else
8560	"ror %[uRet], %[uVal], %[cShift]\n\t"
8561	# endif
8562	: [uRet] "=r" (u32)
8563	: [uVal] "[uRet]" (u32)
8564	, [cShift] "r" (cShift & 31)); /** @todo there is an immediate form here */
8565	return u32;
8566
8567	# else
8568	cShift &= 31;
8569	return (u32 >> cShift) \| (u32 << (32 - cShift));
8570	# endif
8571	}
8572	#endif
8573
8574
8575	/**
8576	* Rotate 64-bit unsigned value to the left by @a cShift.
8577	*
8578	* @returns Rotated value.
8579	* @param u64 The value to rotate.
8580	* @param cShift How many bits to rotate by.
8581	*/
8582	DECLINLINE(uint64_t) ASMRotateLeftU64(uint64_t u64, uint32_t cShift) RT_NOTHROW_DEF
8583	{
8584	#if RT_INLINE_ASM_USES_INTRIN
8585	return _rotl64(u64, cShift);
8586
8587	#elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
8588	__asm__ __volatile__("rolq %b1, %0" : "=g" (u64) : "Jc" (cShift), "0" (u64) : "cc");
8589	return u64;
8590
8591	#elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_X86)
8592	uint32_t uSpill;
8593	__asm__ __volatile__("testb $0x20, %%cl\n\t" /* if (cShift >= 0x20) { swap(u64.hi, u64lo); cShift -= 0x20; } */
8594	"jz 1f\n\t"
8595	"xchgl %%eax, %%edx\n\t"
8596	"1:\n\t"
8597	"andb $0x1f, %%cl\n\t" /* if (cShift & 0x1f) { */
8598	"jz 2f\n\t"
8599	"movl %%edx, %2\n\t" /* save the hi value in %3. */
8600	"shldl %%cl,%%eax,%%edx\n\t" /* shift the hi value left, feeding MSBits from the low value. */
8601	"shldl %%cl,%2,%%eax\n\t" /* shift the lo value left, feeding MSBits from the saved hi value. */
8602	"2:\n\t" /* } */
8603	: "=A" (u64)
8604	, "=c" (cShift)
8605	, "=r" (uSpill)
8606	: "0" (u64)
8607	, "1" (cShift)
8608	: "cc");
8609	return u64;
8610
8611	# elif defined(RT_ARCH_ARM64)
8612	__asm__ __volatile__("ror %[uRet], %[uVal], %[cShift]\n\t"
8613	: [uRet] "=r" (u64)
8614	: [uVal] "[uRet]" (u64)
8615	, [cShift] "r" ((uint64_t)(64 - (cShift & 63)))); /** @todo there is an immediate form here */
8616	return u64;
8617
8618	#else
8619	cShift &= 63;
8620	return (u64 << cShift) \| (u64 >> (64 - cShift));
8621	#endif
8622	}
8623
8624
8625	/**
8626	* Rotate 64-bit unsigned value to the right by @a cShift.
8627	*
8628	* @returns Rotated value.
8629	* @param u64 The value to rotate.
8630	* @param cShift How many bits to rotate by.
8631	*/
8632	DECLINLINE(uint64_t) ASMRotateRightU64(uint64_t u64, uint32_t cShift) RT_NOTHROW_DEF
8633	{
8634	#if RT_INLINE_ASM_USES_INTRIN
8635	return _rotr64(u64, cShift);
8636
8637	#elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64)
8638	__asm__ __volatile__("rorq %b1, %0" : "=g" (u64) : "Jc" (cShift), "0" (u64) : "cc");
8639	return u64;
8640
8641	#elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_X86)
8642	uint32_t uSpill;
8643	__asm__ __volatile__("testb $0x20, %%cl\n\t" /* if (cShift >= 0x20) { swap(u64.hi, u64lo); cShift -= 0x20; } */
8644	"jz 1f\n\t"
8645	"xchgl %%eax, %%edx\n\t"
8646	"1:\n\t"
8647	"andb $0x1f, %%cl\n\t" /* if (cShift & 0x1f) { */
8648	"jz 2f\n\t"
8649	"movl %%edx, %2\n\t" /* save the hi value in %3. */
8650	"shrdl %%cl,%%eax,%%edx\n\t" /* shift the hi value right, feeding LSBits from the low value. */
8651	"shrdl %%cl,%2,%%eax\n\t" /* shift the lo value right, feeding LSBits from the saved hi value. */
8652	"2:\n\t" /* } */
8653	: "=A" (u64)
8654	, "=c" (cShift)
8655	, "=r" (uSpill)
8656	: "0" (u64)
8657	, "1" (cShift)
8658	: "cc");
8659	return u64;
8660
8661	# elif defined(RT_ARCH_ARM64)
8662	__asm__ __volatile__("ror %[uRet], %[uVal], %[cShift]\n\t"
8663	: [uRet] "=r" (u64)
8664	: [uVal] "[uRet]" (u64)
8665	, [cShift] "r" ((uint64_t)(cShift & 63))); /** @todo there is an immediate form here */
8666	return u64;
8667
8668	#else
8669	cShift &= 63;
8670	return (u64 >> cShift) \| (u64 << (64 - cShift));
8671	#endif
8672	}
8673
8674	/** @} */
8675
8676
8677	/** @} */
8678
8679	/*
8680	* Include #pragma aux definitions for Watcom C/C++.
8681	*/
8682	#if defined(__WATCOMC__) && ARCH_BITS == 16 && defined(RT_ARCH_X86)
8683	# define IPRT_ASM_WATCOM_X86_16_WITH_PRAGMAS
8684	# undef IPRT_INCLUDED_asm_watcom_x86_16_h
8685	# include "asm-watcom-x86-16.h"
8686	#elif defined(__WATCOMC__) && ARCH_BITS == 32 && defined(RT_ARCH_X86)
8687	# define IPRT_ASM_WATCOM_X86_32_WITH_PRAGMAS
8688	# undef IPRT_INCLUDED_asm_watcom_x86_32_h
8689	# include "asm-watcom-x86-32.h"
8690	#endif
8691
8692	#endif /* !IPRT_INCLUDED_asm_h */
8693

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

source: vbox/trunk/include/iprt/asm.h@ 105688

Download in other formats: