tstRTInlineAsm.cpp@ 93755

Last change on this file since 93755 was 93755, checked in by vboxsync, 3 years ago
tstRTInlineAsm: Use picoseconds instead of nanoseconds when benchmarking inline assembly operations. bugref:9898
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 151.2 KB

Line
1	/* $Id: tstRTInlineAsm.cpp 93755 2022-02-15 14:32:55Z vboxsync $ */
2	/** @file
3	* IPRT Testcase - inline assembly.
4	*/
5
6	/*
7	* Copyright (C) 2006-2022 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27
28	/*********************************************************************************************************************************
29	* Header Files *
30	*********************************************************************************************************************************/
31	#include <iprt/asm.h>
32	#include <iprt/asm-math.h>
33
34	/* See http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44018. Only gcc version 4.4
35	* is affected. No harm for the VBox code: If the cpuid code compiles, it works
36	* fine. */
37	#if defined(__GNUC__) && defined(RT_ARCH_X86) && defined(__PIC__)
38	# if __GNUC__ == 4 && __GNUC_MINOR__ == 4
39	# define GCC44_32BIT_PIC
40	# endif
41	#endif
42
43	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
44	# include <iprt/asm-amd64-x86.h>
45	# include <iprt/x86.h>
46	#elif defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)
47	# include <iprt/asm-arm.h>
48	# include <iprt/time.h>
49	#else
50	# include <iprt/time.h>
51	#endif
52	#include <iprt/mem.h>
53	#include <iprt/param.h>
54	#include <iprt/rand.h>
55	#include <iprt/stream.h>
56	#include <iprt/string.h>
57	#include <iprt/thread.h>
58	#include <iprt/test.h>
59	#include <iprt/time.h>
60
61
62
63	/*********************************************************************************************************************************
64	* Defined Constants And Macros *
65	*********************************************************************************************************************************/
66	#define CHECKVAL(val, expect, fmt) \
67	do \
68	{ \
69	if ((val) != (expect)) \
70	{ \
71	RTTestFailed(g_hTest, "%s, %d: " #val ": expected " fmt " got " fmt "\n", __FUNCTION__, __LINE__, (expect), (val)); \
72	} \
73	} while (0)
74
75	#define CHECKOP(op, expect, fmt, type) \
76	do \
77	{ \
78	type val = op; \
79	if (val != (type)(expect)) \
80	{ \
81	RTTestFailed(g_hTest, "%s, %d: " #op ": expected " fmt " got " fmt "\n", __FUNCTION__, __LINE__, (type)(expect), val); \
82	} \
83	} while (0)
84
85	#define CHECK_OP_AND_VAL(a_Type, a_Fmt, a_pVar, a_Operation, a_ExpectRetVal, a_ExpectVarVal) \
86	do { \
87	CHECKOP(a_Operation, a_ExpectRetVal, a_Fmt, a_Type); \
88	CHECKVAL(*a_pVar, a_ExpectVarVal, a_Fmt); \
89	} while (0)
90
91	#define CHECK_OP_AND_VAL_EX(a_TypeRet, a_FmtRet, a_FmtVar, a_pVar, a_Operation, a_ExpectRetVal, a_ExpectVarVal) \
92	do { \
93	CHECKOP(a_Operation, a_ExpectRetVal, a_FmtRet, a_TypeRet); \
94	CHECKVAL(*a_pVar, a_ExpectVarVal, a_FmtVar); \
95	} while (0)
96
97	#define CHECK_OP_AND_VAL_EX2(a_TypeRet, a_FmtRet, a_FmtVar, a_pVar, a_uVar2, a_Operation, a_ExpectRetVal, a_ExpectVarVal, a_ExpectVarVal2) \
98	do { \
99	CHECKOP(a_Operation, a_ExpectRetVal, a_FmtRet, a_TypeRet); \
100	CHECKVAL(*a_pVar, a_ExpectVarVal, a_FmtVar); \
101	CHECKVAL(a_uVar2, a_ExpectVarVal2, a_FmtVar); \
102	} while (0)
103
104	#define CHECKVAL128(a_pu128Val, a_u64HiExpect, a_u64LoExpect) \
105	do \
106	{ \
107	if ((a_pu128Val)->s.Hi != (a_u64HiExpect) \|\| (a_pu128Val)->s.Lo != (a_u64LoExpect)) \
108	RTTestFailed(g_hTest, "%s, %d: " #a_pu128Val ": expected %#RX64'%016RX64 got %#RX64'%016RX64\n", \
109	__FUNCTION__, __LINE__, (a_u64HiExpect), (a_u64LoExpect), (a_pu128Val)->s.Hi, (a_pu128Val)->s.Lo); \
110	} while (0)
111	#define CHECKVAL128_C(a_pu128Val, a_u64HiExpect, a_u64LoExpect) \
112	do \
113	{ \
114	if ((a_pu128Val)->s.Hi != UINT64_C(a_u64HiExpect) \|\| (a_pu128Val)->s.Lo != UINT64_C(a_u64LoExpect)) \
115	RTTestFailed(g_hTest, "%s, %d: " #a_pu128Val ": expected %#RX64'%016RX64 got %#RX64'%016RX64\n", \
116	__FUNCTION__, __LINE__, UINT64_C(a_u64HiExpect), UINT64_C(a_u64LoExpect), \
117	(a_pu128Val)->s.Hi, (a_pu128Val)->s.Lo); \
118	} while (0)
119	#define CHECK_OP_AND_VAL_128(a_TypeRet, a_FmtRet, a_pu128Val, a_Operation, a_ExpectRetVal, a_u64HiExpect, a_u64LoExpect) \
120	do { \
121	CHECKOP(a_Operation, a_ExpectRetVal, a_FmtRet, a_TypeRet); \
122	CHECKVAL128(a_pu128Val, a_u64HiExpect, a_u64LoExpect); \
123	} while (0)
124	#define CHECK_OP_AND_VAL_128_C(a_TypeRet, a_FmtRet, a_pu128Val, a_Operation, a_ExpectRetVal, a_u64HiExpect, a_u64LoExpect) \
125	do { \
126	CHECKOP(a_Operation, a_ExpectRetVal, a_FmtRet, a_TypeRet); \
127	CHECKVAL128_C(a_pu128Val, a_u64HiExpect, a_u64LoExpect); \
128	} while (0)
129
130	/**
131	* Calls a worker function with different worker variable storage types.
132	*/
133	#define DO_SIMPLE_TEST_NO_SUB_NO_STACK(a_WorkerFunction, type) \
134	do \
135	{ \
136	type pVar = (type )RTTestGuardedAllocHead(g_hTest, sizeof(type)); \
137	RTTEST_CHECK_BREAK(g_hTest, pVar); \
138	a_WorkerFunction(pVar); \
139	RTTestGuardedFree(g_hTest, pVar); \
140	\
141	pVar = (type *)RTTestGuardedAllocTail(g_hTest, sizeof(type)); \
142	RTTEST_CHECK_BREAK(g_hTest, pVar); \
143	a_WorkerFunction(pVar); \
144	RTTestGuardedFree(g_hTest, pVar); \
145	} while (0)
146
147
148	/**
149	* Calls a worker function with different worker variable storage types.
150	*/
151	#define DO_SIMPLE_TEST_NO_SUB(a_WorkerFunction, type) \
152	do \
153	{ \
154	type StackVar; \
155	a_WorkerFunction(&StackVar); \
156	DO_SIMPLE_TEST_NO_SUB_NO_STACK(a_WorkerFunction, type); \
157	} while (0)
158
159	/**
160	* Calls a worker function with different worker variable storage types.
161	*/
162	#define DO_SIMPLE_TEST(name, type) \
163	do \
164	{ \
165	RTTestISub(#name); \
166	DO_SIMPLE_TEST_NO_SUB(tst ## name ## Worker, type); \
167	} while (0)
168
169
170	/*********************************************************************************************************************************
171	* Global Variables *
172	*********************************************************************************************************************************/
173	/** The test instance. */
174	static RTTEST g_hTest;
175
176
177
178	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
179
180	const char *getCacheAss(unsigned u)
181	{
182	if (u == 0)
183	return "res0 ";
184	if (u == 1)
185	return "direct";
186	if (u >= 256)
187	return "???";
188
189	char *pszRet = NULL;
190	RTStrAPrintf(&pszRet, "%d way", u);
191	RTMEM_WILL_LEAK(pszRet);
192	return pszRet;
193	}
194
195
196	const char *getL2CacheAss(unsigned u)
197	{
198	switch (u)
199	{
200	case 0: return "off ";
201	case 1: return "direct";
202	case 2: return "2 way ";
203	case 3: return "res3 ";
204	case 4: return "4 way ";
205	case 5: return "res5 ";
206	case 6: return "8 way ";
207	case 7: return "res7 ";
208	case 8: return "16 way";
209	case 9: return "res9 ";
210	case 10: return "res10 ";
211	case 11: return "res11 ";
212	case 12: return "res12 ";
213	case 13: return "res13 ";
214	case 14: return "res14 ";
215	case 15: return "fully ";
216	default:
217	return "????";
218	}
219	}
220
221
222	/**
223	* Test and dump all possible info from the CPUID instruction.
224	*
225	* @remark Bits shared with the libc cpuid.c program. This all written by me, so no worries.
226	* @todo transform the dumping into a generic runtime function. We'll need it for logging!
227	*/
228	void tstASMCpuId(void)
229	{
230	RTTestISub("ASMCpuId");
231
232	unsigned iBit;
233	struct
234	{
235	uint32_t uEBX, uEAX, uEDX, uECX;
236	} s;
237	if (!ASMHasCpuId())
238	{
239	RTTestIPrintf(RTTESTLVL_ALWAYS, "warning! CPU doesn't support CPUID\n");
240	return;
241	}
242
243	/*
244	* Try the 0 function and use that for checking the ASMCpuId_* variants.
245	*/
246	ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
247
248	uint32_t u32;
249
250	u32 = ASMCpuId_EAX(0);
251	CHECKVAL(u32, s.uEAX, "%x");
252	u32 = ASMCpuId_EBX(0);
253	CHECKVAL(u32, s.uEBX, "%x");
254	u32 = ASMCpuId_ECX(0);
255	CHECKVAL(u32, s.uECX, "%x");
256	u32 = ASMCpuId_EDX(0);
257	CHECKVAL(u32, s.uEDX, "%x");
258
259	uint32_t uECX2 = s.uECX - 1;
260	uint32_t uEDX2 = s.uEDX - 1;
261	ASMCpuId_ECX_EDX(0, &uECX2, &uEDX2);
262	CHECKVAL(uECX2, s.uECX, "%x");
263	CHECKVAL(uEDX2, s.uEDX, "%x");
264
265	uint32_t uEAX2 = s.uEAX - 1;
266	uint32_t uEBX2 = s.uEBX - 1;
267	uECX2 = s.uECX - 1;
268	uEDX2 = s.uEDX - 1;
269	ASMCpuIdExSlow(0, 0, 0, 0, &uEAX2, &uEBX2, &uECX2, &uEDX2);
270	CHECKVAL(uEAX2, s.uEAX, "%x");
271	CHECKVAL(uEBX2, s.uEBX, "%x");
272	CHECKVAL(uECX2, s.uECX, "%x");
273	CHECKVAL(uEDX2, s.uEDX, "%x");
274
275	/*
276	* Check the extended APIC stuff.
277	*/
278	uint32_t idExtApic;
279	if (ASMCpuId_EAX(0) >= 0xb)
280	{
281	uint8_t idApic = ASMGetApicId();
282	do
283	{
284	uEAX2 = uEBX2 = uECX2 = uEDX2 = UINT32_C(0x50486744);
285	ASMCpuIdExSlow(0xb, 0, 0, 0, &uEAX2, &uEBX2, &uECX2, &uEDX2);
286	idExtApic = ASMGetApicIdExt0B();
287	} while (ASMGetApicId() != idApic);
288
289	CHECKVAL(uEDX2, idExtApic, "%x");
290	if (idApic != (uint8_t)idExtApic && uECX2 != 0)
291	RTTestIFailed("ASMGetApicIdExt0B() -> %#x vs ASMGetApicId() -> %#x", idExtApic, idApic);
292	}
293	if (ASMCpuId_EAX(UINT32_C(0x80000000)) >= UINT32_C(0x8000001E))
294	{
295	uint8_t idApic = ASMGetApicId();
296	do
297	{
298	uEAX2 = uEBX2 = uECX2 = uEDX2 = UINT32_C(0x50486744);
299	ASMCpuIdExSlow(0x8000001e, 0, 0, 0, &uEAX2, &uEBX2, &uECX2, &uEDX2);
300	idExtApic = ASMGetApicIdExt8000001E();
301	} while (ASMGetApicId() != idApic);
302	CHECKVAL(uEAX2, idExtApic, "%x");
303	if (idApic != (uint8_t)idExtApic)
304	RTTestIFailed("ASMGetApicIdExt8000001E() -> %#x vs ASMGetApicId() -> %#x", idExtApic, idApic);
305	}
306
307	/*
308	* Done testing, dump the information.
309	*/
310	RTTestIPrintf(RTTESTLVL_ALWAYS, "CPUID Dump\n");
311	ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
312	const uint32_t cFunctions = s.uEAX;
313
314	/* raw dump */
315	RTTestIPrintf(RTTESTLVL_ALWAYS,
316	"\n"
317	" RAW Standard CPUIDs\n"
318	"Function eax ebx ecx edx\n");
319	for (unsigned iStd = 0; iStd <= cFunctions + 3; iStd++)
320	{
321	ASMCpuId_Idx_ECX(iStd, 0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
322	RTTestIPrintf(RTTESTLVL_ALWAYS, "%08x %08x %08x %08x %08x%s\n",
323	iStd, s.uEAX, s.uEBX, s.uECX, s.uEDX, iStd <= cFunctions ? "" : "*");
324
325	/* Some leafs output depend on the initial value of ECX.
326	* The same seems to apply to invalid standard functions */
327	if (iStd > cFunctions)
328	continue;
329	if (iStd == 0x04) /* Deterministic Cache Parameters Leaf */
330	for (uint32_t uECX = 1; s.uEAX & 0x1f; uECX++)
331	{
332	ASMCpuId_Idx_ECX(iStd, uECX, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
333	RTTestIPrintf(RTTESTLVL_ALWAYS, " [%02x] %08x %08x %08x %08x\n", uECX, s.uEAX, s.uEBX, s.uECX, s.uEDX);
334	RTTESTI_CHECK_BREAK(uECX < 128);
335	}
336	else if (iStd == 0x07) /* Structured Extended Feature Flags */
337	{
338	uint32_t uMax = s.uEAX;
339	for (uint32_t uECX = 1; uECX < uMax; uECX++)
340	{
341	ASMCpuId_Idx_ECX(iStd, uECX, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
342	RTTestIPrintf(RTTESTLVL_ALWAYS, " [%02x] %08x %08x %08x %08x\n", uECX, s.uEAX, s.uEBX, s.uECX, s.uEDX);
343	RTTESTI_CHECK_BREAK(uECX < 128);
344	}
345	}
346	else if (iStd == 0x0b) /* Extended Topology Enumeration Leafs */
347	for (uint32_t uECX = 1; (s.uEAX & 0x1f) && (s.uEBX & 0xffff); uECX++)
348	{
349	ASMCpuId_Idx_ECX(iStd, uECX, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
350	RTTestIPrintf(RTTESTLVL_ALWAYS, " [%02x] %08x %08x %08x %08x\n", uECX, s.uEAX, s.uEBX, s.uECX, s.uEDX);
351	RTTESTI_CHECK_BREAK(uECX < 128);
352	}
353	else if (iStd == 0x0d) /* Extended State Enumeration Leafs */
354	for (uint32_t uECX = 1; s.uEAX != 0 \|\| s.uEBX != 0 \|\| s.uECX != 0 \|\| s.uEDX != 0; uECX++)
355	{
356	ASMCpuId_Idx_ECX(iStd, uECX, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
357	RTTestIPrintf(RTTESTLVL_ALWAYS, " [%02x] %08x %08x %08x %08x\n", uECX, s.uEAX, s.uEBX, s.uECX, s.uEDX);
358	RTTESTI_CHECK_BREAK(uECX < 128);
359	}
360	else if ( iStd == 0x0f /* Platform quality of service monitoring (PQM) */
361	\|\| iStd == 0x10 /* Platform quality of service enforcement (PQE) */
362	\|\| iStd == 0x12 /* SGX Enumeration */
363	\|\| iStd == 0x14 /* Processor Trace Enumeration */
364	\|\| iStd == 0x17 /* SoC Vendor Attribute Enumeration */
365	\|\| iStd == 0x18 /* Deterministic Address Translation Parameters */)
366	{
367	/** @todo */
368	}
369	else
370	{
371	u32 = ASMCpuId_EAX(iStd);
372	CHECKVAL(u32, s.uEAX, "%x");
373
374	uint32_t u32EbxMask = UINT32_MAX;
375	if (iStd == 1)
376	u32EbxMask = UINT32_C(0x00ffffff); /* Omit the local apic ID in case we're rescheduled. */
377	u32 = ASMCpuId_EBX(iStd);
378	CHECKVAL(u32 & u32EbxMask, s.uEBX & u32EbxMask, "%x");
379
380	u32 = ASMCpuId_ECX(iStd);
381	CHECKVAL(u32, s.uECX, "%x");
382	u32 = ASMCpuId_EDX(iStd);
383	CHECKVAL(u32, s.uEDX, "%x");
384
385	uECX2 = s.uECX - 1;
386	uEDX2 = s.uEDX - 1;
387	ASMCpuId_ECX_EDX(iStd, &uECX2, &uEDX2);
388	CHECKVAL(uECX2, s.uECX, "%x");
389	CHECKVAL(uEDX2, s.uEDX, "%x");
390
391	uEAX2 = s.uEAX - 1;
392	uEBX2 = s.uEBX - 1;
393	uECX2 = s.uECX - 1;
394	uEDX2 = s.uEDX - 1;
395	ASMCpuId(iStd, &uEAX2, &uEBX2, &uECX2, &uEDX2);
396	CHECKVAL(uEAX2, s.uEAX, "%x");
397	CHECKVAL(uEBX2 & u32EbxMask, s.uEBX & u32EbxMask, "%x");
398	CHECKVAL(uECX2, s.uECX, "%x");
399	CHECKVAL(uEDX2, s.uEDX, "%x");
400	}
401	}
402
403	/*
404	* Understandable output
405	*/
406	ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
407	RTTestIPrintf(RTTESTLVL_ALWAYS,
408	"Name: %.04s%.04s%.04s\n"
409	"Support: 0-%u\n",
410	&s.uEBX, &s.uEDX, &s.uECX, s.uEAX);
411	bool const fIntel = RTX86IsIntelCpu(s.uEBX, s.uECX, s.uEDX);
412
413	/*
414	* Get Features.
415	*/
416	if (cFunctions >= 1)
417	{
418	static const char * const s_apszTypes[4] = { "primary", "overdrive", "MP", "reserved" };
419	ASMCpuId(1, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
420	RTTestIPrintf(RTTESTLVL_ALWAYS,
421	"Family: %#x \tExtended: %#x \tEffective: %#x\n"
422	"Model: %#x \tExtended: %#x \tEffective: %#x\n"
423	"Stepping: %d\n"
424	"Type: %d (%s)\n"
425	"APIC ID: %#04x\n"
426	"Logical CPUs: %d\n"
427	"CLFLUSH Size: %d\n"
428	"Brand ID: %#04x\n",
429	(s.uEAX >> 8) & 0xf, (s.uEAX >> 20) & 0x7f, RTX86GetCpuFamily(s.uEAX),
430	(s.uEAX >> 4) & 0xf, (s.uEAX >> 16) & 0x0f, RTX86GetCpuModel(s.uEAX, fIntel),
431	RTX86GetCpuStepping(s.uEAX),
432	(s.uEAX >> 12) & 0x3, s_apszTypes[(s.uEAX >> 12) & 0x3],
433	(s.uEBX >> 24) & 0xff,
434	(s.uEBX >> 16) & 0xff,
435	(s.uEBX >> 8) & 0xff,
436	(s.uEBX >> 0) & 0xff);
437
438	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features EDX: ");
439	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FPU");
440	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VME");
441	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DE");
442	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE");
443	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TSC");
444	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MSR");
445	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAE");
446	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCE");
447	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CX8");
448	if (s.uEDX & RT_BIT(9)) RTTestIPrintf(RTTESTLVL_ALWAYS, " APIC");
449	if (s.uEDX & RT_BIT(10)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 10");
450	if (s.uEDX & RT_BIT(11)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SEP");
451	if (s.uEDX & RT_BIT(12)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MTRR");
452	if (s.uEDX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PGE");
453	if (s.uEDX & RT_BIT(14)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCA");
454	if (s.uEDX & RT_BIT(15)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMOV");
455	if (s.uEDX & RT_BIT(16)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAT");
456	if (s.uEDX & RT_BIT(17)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE36");
457	if (s.uEDX & RT_BIT(18)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSN");
458	if (s.uEDX & RT_BIT(19)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CLFSH");
459	if (s.uEDX & RT_BIT(20)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 20");
460	if (s.uEDX & RT_BIT(21)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DS");
461	if (s.uEDX & RT_BIT(22)) RTTestIPrintf(RTTESTLVL_ALWAYS, " ACPI");
462	if (s.uEDX & RT_BIT(23)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MMX");
463	if (s.uEDX & RT_BIT(24)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FXSR");
464	if (s.uEDX & RT_BIT(25)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE");
465	if (s.uEDX & RT_BIT(26)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE2");
466	if (s.uEDX & RT_BIT(27)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SS");
467	if (s.uEDX & RT_BIT(28)) RTTestIPrintf(RTTESTLVL_ALWAYS, " HTT");
468	if (s.uEDX & RT_BIT(29)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 29");
469	if (s.uEDX & RT_BIT(30)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 30");
470	if (s.uEDX & RT_BIT(31)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 31");
471	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
472
473	/** @todo check intel docs. */
474	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features ECX: ");
475	if (s.uECX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE3");
476	for (iBit = 1; iBit < 13; iBit++)
477	if (s.uECX & RT_BIT(iBit))
478	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
479	if (s.uECX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CX16");
480	for (iBit = 14; iBit < 32; iBit++)
481	if (s.uECX & RT_BIT(iBit))
482	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
483	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
484	}
485	if (ASMCpuId_EAX(0) >= 0xb)
486	RTTestIPrintf(RTTESTLVL_ALWAYS, "APIC ID(Ext 0b): %#010x\n", ASMGetApicIdExt0B());
487
488	/*
489	* Extended.
490	* Implemented after AMD specs.
491	*/
492	/** @todo check out the intel specs. */
493	ASMCpuId(0x80000000, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
494	if (!s.uEAX && !s.uEBX && !s.uECX && !s.uEDX)
495	{
496	RTTestIPrintf(RTTESTLVL_ALWAYS, "No extended CPUID info? Check the manual on how to detect this...\n");
497	return;
498	}
499	const uint32_t cExtFunctions = s.uEAX \| 0x80000000;
500
501	/* raw dump */
502	RTTestIPrintf(RTTESTLVL_ALWAYS,
503	"\n"
504	" RAW Extended CPUIDs\n"
505	"Function eax ebx ecx edx\n");
506	for (unsigned iExt = 0x80000000; iExt <= cExtFunctions + 3; iExt++)
507	{
508	ASMCpuId(iExt, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
509	RTTestIPrintf(RTTESTLVL_ALWAYS, "%08x %08x %08x %08x %08x%s\n",
510	iExt, s.uEAX, s.uEBX, s.uECX, s.uEDX, iExt <= cExtFunctions ? "" : "*");
511
512	if (iExt > cExtFunctions)
513	continue; /* Invalid extended functions seems change the value if ECX changes */
514	if (iExt == 0x8000001d)
515	continue; /* Takes cache level in ecx. */
516
517	u32 = ASMCpuId_EAX(iExt);
518	CHECKVAL(u32, s.uEAX, "%x");
519	u32 = ASMCpuId_EBX(iExt);
520	CHECKVAL(u32, s.uEBX, "%x");
521	u32 = ASMCpuId_ECX(iExt);
522	CHECKVAL(u32, s.uECX, "%x");
523	u32 = ASMCpuId_EDX(iExt);
524	CHECKVAL(u32, s.uEDX, "%x");
525
526	uECX2 = s.uECX - 1;
527	uEDX2 = s.uEDX - 1;
528	ASMCpuId_ECX_EDX(iExt, &uECX2, &uEDX2);
529	CHECKVAL(uECX2, s.uECX, "%x");
530	CHECKVAL(uEDX2, s.uEDX, "%x");
531
532	uEAX2 = s.uEAX - 1;
533	uEBX2 = s.uEBX - 1;
534	uECX2 = s.uECX - 1;
535	uEDX2 = s.uEDX - 1;
536	ASMCpuId(iExt, &uEAX2, &uEBX2, &uECX2, &uEDX2);
537	CHECKVAL(uEAX2, s.uEAX, "%x");
538	CHECKVAL(uEBX2, s.uEBX, "%x");
539	CHECKVAL(uECX2, s.uECX, "%x");
540	CHECKVAL(uEDX2, s.uEDX, "%x");
541	}
542
543	/*
544	* Understandable output
545	*/
546	ASMCpuId(0x80000000, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
547	RTTestIPrintf(RTTESTLVL_ALWAYS,
548	"Ext Name: %.4s%.4s%.4s\n"
549	"Ext Supports: 0x80000000-%#010x\n",
550	&s.uEBX, &s.uEDX, &s.uECX, s.uEAX);
551
552	if (cExtFunctions >= 0x80000001)
553	{
554	ASMCpuId(0x80000001, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
555	RTTestIPrintf(RTTESTLVL_ALWAYS,
556	"Family: %#x \tExtended: %#x \tEffective: %#x\n"
557	"Model: %#x \tExtended: %#x \tEffective: %#x\n"
558	"Stepping: %d\n"
559	"Brand ID: %#05x\n",
560	(s.uEAX >> 8) & 0xf, (s.uEAX >> 20) & 0x7f, RTX86GetCpuFamily(s.uEAX),
561	(s.uEAX >> 4) & 0xf, (s.uEAX >> 16) & 0x0f, RTX86GetCpuModel(s.uEAX, fIntel),
562	RTX86GetCpuStepping(s.uEAX),
563	s.uEBX & 0xfff);
564
565	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features EDX: ");
566	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FPU");
567	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VME");
568	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DE");
569	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE");
570	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TSC");
571	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MSR");
572	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAE");
573	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCE");
574	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMPXCHG8B");
575	if (s.uEDX & RT_BIT(9)) RTTestIPrintf(RTTESTLVL_ALWAYS, " APIC");
576	if (s.uEDX & RT_BIT(10)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 10");
577	if (s.uEDX & RT_BIT(11)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SysCallSysRet");
578	if (s.uEDX & RT_BIT(12)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MTRR");
579	if (s.uEDX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PGE");
580	if (s.uEDX & RT_BIT(14)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCA");
581	if (s.uEDX & RT_BIT(15)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMOV");
582	if (s.uEDX & RT_BIT(16)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAT");
583	if (s.uEDX & RT_BIT(17)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE36");
584	if (s.uEDX & RT_BIT(18)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 18");
585	if (s.uEDX & RT_BIT(19)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 19");
586	if (s.uEDX & RT_BIT(20)) RTTestIPrintf(RTTESTLVL_ALWAYS, " NX");
587	if (s.uEDX & RT_BIT(21)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 21");
588	if (s.uEDX & RT_BIT(22)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MmxExt");
589	if (s.uEDX & RT_BIT(23)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MMX");
590	if (s.uEDX & RT_BIT(24)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FXSR");
591	if (s.uEDX & RT_BIT(25)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FastFXSR");
592	if (s.uEDX & RT_BIT(26)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 26");
593	if (s.uEDX & RT_BIT(27)) RTTestIPrintf(RTTESTLVL_ALWAYS, " RDTSCP");
594	if (s.uEDX & RT_BIT(28)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 28");
595	if (s.uEDX & RT_BIT(29)) RTTestIPrintf(RTTESTLVL_ALWAYS, " LongMode");
596	if (s.uEDX & RT_BIT(30)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3DNowExt");
597	if (s.uEDX & RT_BIT(31)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3DNow");
598	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
599
600	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features ECX: ");
601	if (s.uECX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " LahfSahf");
602	if (s.uECX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CmpLegacy");
603	if (s.uECX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SVM");
604	if (s.uECX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3");
605	if (s.uECX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " AltMovCr8");
606	for (iBit = 5; iBit < 32; iBit++)
607	if (s.uECX & RT_BIT(iBit))
608	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
609	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
610	}
611
612	char szString[443+1] = {0};
613	if (cExtFunctions >= 0x80000002)
614	ASMCpuId(0x80000002, &szString[0 + 0], &szString[0 + 4], &szString[0 + 8], &szString[0 + 12]);
615	if (cExtFunctions >= 0x80000003)
616	ASMCpuId(0x80000003, &szString[16 + 0], &szString[16 + 4], &szString[16 + 8], &szString[16 + 12]);
617	if (cExtFunctions >= 0x80000004)
618	ASMCpuId(0x80000004, &szString[32 + 0], &szString[32 + 4], &szString[32 + 8], &szString[32 + 12]);
619	if (cExtFunctions >= 0x80000002)
620	RTTestIPrintf(RTTESTLVL_ALWAYS, "Full Name: %s\n", szString);
621
622	if (cExtFunctions >= 0x80000005)
623	{
624	ASMCpuId(0x80000005, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
625	RTTestIPrintf(RTTESTLVL_ALWAYS,
626	"TLB 2/4M Instr/Uni: %s %3d entries\n"
627	"TLB 2/4M Data: %s %3d entries\n",
628	getCacheAss((s.uEAX >> 8) & 0xff), (s.uEAX >> 0) & 0xff,
629	getCacheAss((s.uEAX >> 24) & 0xff), (s.uEAX >> 16) & 0xff);
630	RTTestIPrintf(RTTESTLVL_ALWAYS,
631	"TLB 4K Instr/Uni: %s %3d entries\n"
632	"TLB 4K Data: %s %3d entries\n",
633	getCacheAss((s.uEBX >> 8) & 0xff), (s.uEBX >> 0) & 0xff,
634	getCacheAss((s.uEBX >> 24) & 0xff), (s.uEBX >> 16) & 0xff);
635	RTTestIPrintf(RTTESTLVL_ALWAYS,
636	"L1 Instr Cache Line Size: %d bytes\n"
637	"L1 Instr Cache Lines Per Tag: %d\n"
638	"L1 Instr Cache Associativity: %s\n"
639	"L1 Instr Cache Size: %d KB\n",
640	(s.uEDX >> 0) & 0xff,
641	(s.uEDX >> 8) & 0xff,
642	getCacheAss((s.uEDX >> 16) & 0xff),
643	(s.uEDX >> 24) & 0xff);
644	RTTestIPrintf(RTTESTLVL_ALWAYS,
645	"L1 Data Cache Line Size: %d bytes\n"
646	"L1 Data Cache Lines Per Tag: %d\n"
647	"L1 Data Cache Associativity: %s\n"
648	"L1 Data Cache Size: %d KB\n",
649	(s.uECX >> 0) & 0xff,
650	(s.uECX >> 8) & 0xff,
651	getCacheAss((s.uECX >> 16) & 0xff),
652	(s.uECX >> 24) & 0xff);
653	}
654
655	if (cExtFunctions >= 0x80000006)
656	{
657	ASMCpuId(0x80000006, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
658	RTTestIPrintf(RTTESTLVL_ALWAYS,
659	"L2 TLB 2/4M Instr/Uni: %s %4d entries\n"
660	"L2 TLB 2/4M Data: %s %4d entries\n",
661	getL2CacheAss((s.uEAX >> 12) & 0xf), (s.uEAX >> 0) & 0xfff,
662	getL2CacheAss((s.uEAX >> 28) & 0xf), (s.uEAX >> 16) & 0xfff);
663	RTTestIPrintf(RTTESTLVL_ALWAYS,
664	"L2 TLB 4K Instr/Uni: %s %4d entries\n"
665	"L2 TLB 4K Data: %s %4d entries\n",
666	getL2CacheAss((s.uEBX >> 12) & 0xf), (s.uEBX >> 0) & 0xfff,
667	getL2CacheAss((s.uEBX >> 28) & 0xf), (s.uEBX >> 16) & 0xfff);
668	RTTestIPrintf(RTTESTLVL_ALWAYS,
669	"L2 Cache Line Size: %d bytes\n"
670	"L2 Cache Lines Per Tag: %d\n"
671	"L2 Cache Associativity: %s\n"
672	"L2 Cache Size: %d KB\n",
673	(s.uEDX >> 0) & 0xff,
674	(s.uEDX >> 8) & 0xf,
675	getL2CacheAss((s.uEDX >> 12) & 0xf),
676	(s.uEDX >> 16) & 0xffff);
677	}
678
679	if (cExtFunctions >= 0x80000007)
680	{
681	ASMCpuId(0x80000007, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
682	RTTestIPrintf(RTTESTLVL_ALWAYS, "APM Features: ");
683	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TS");
684	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FID");
685	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VID");
686	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TTP");
687	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TM");
688	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " STC");
689	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 6");
690	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 7");
691	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TscInvariant");
692	for (iBit = 9; iBit < 32; iBit++)
693	if (s.uEDX & RT_BIT(iBit))
694	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
695	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
696	}
697
698	if (cExtFunctions >= 0x80000008)
699	{
700	ASMCpuId(0x80000008, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
701	RTTestIPrintf(RTTESTLVL_ALWAYS,
702	"Physical Address Width: %d bits\n"
703	"Virtual Address Width: %d bits\n"
704	"Guest Physical Address Width: %d bits\n",
705	(s.uEAX >> 0) & 0xff,
706	(s.uEAX >> 8) & 0xff,
707	(s.uEAX >> 16) & 0xff);
708	RTTestIPrintf(RTTESTLVL_ALWAYS,
709	"Physical Core Count: %d\n",
710	((s.uECX >> 0) & 0xff) + 1);
711	if ((s.uECX >> 12) & 0xf)
712	RTTestIPrintf(RTTESTLVL_ALWAYS, "ApicIdCoreIdSize: %d bits\n", (s.uECX >> 12) & 0xf);
713	}
714
715	if (cExtFunctions >= 0x8000000a)
716	{
717	ASMCpuId(0x8000000a, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
718	RTTestIPrintf(RTTESTLVL_ALWAYS,
719	"SVM Revision: %d (%#x)\n"
720	"Number of Address Space IDs: %d (%#x)\n",
721	s.uEAX & 0xff, s.uEAX & 0xff,
722	s.uEBX, s.uEBX);
723	}
724	if (ASMCpuId_EAX(UINT32_C(0x80000000)) >= UINT32_C(0x8000001E))
725	RTTestIPrintf(RTTESTLVL_ALWAYS, "APIC ID(Ext 8000001b): %#010x\n", ASMGetApicIdExt8000001E());
726	}
727
728	# if 0
729	static void bruteForceCpuId(void)
730	{
731	RTTestISub("brute force CPUID leafs");
732	uint32_t auPrevValues[4] = { 0, 0, 0, 0};
733	uint32_t uLeaf = 0;
734	do
735	{
736	uint32_t auValues[4];
737	ASMCpuIdExSlow(uLeaf, 0, 0, 0, &auValues[0], &auValues[1], &auValues[2], &auValues[3]);
738	if ( (auValues[0] != auPrevValues[0] && auValues[0] != uLeaf)
739	\|\| (auValues[1] != auPrevValues[1] && auValues[1] != 0)
740	\|\| (auValues[2] != auPrevValues[2] && auValues[2] != 0)
741	\|\| (auValues[3] != auPrevValues[3] && auValues[3] != 0)
742	\|\| (uLeaf & (UINT32_C(0x08000000) - UINT32_C(1))) == 0)
743	{
744	RTTestIPrintf(RTTESTLVL_ALWAYS,
745	"%08x: %08x %08x %08x %08x\n", uLeaf,
746	auValues[0], auValues[1], auValues[2], auValues[3]);
747	}
748	auPrevValues[0] = auValues[0];
749	auPrevValues[1] = auValues[1];
750	auPrevValues[2] = auValues[2];
751	auPrevValues[3] = auValues[3];
752
753	//uint32_t uSubLeaf = 0;
754	//do
755	//{
756	//
757	//
758	//} while (false);
759	} while (uLeaf++ < UINT32_MAX);
760	}
761	# endif
762
763	#endif /* AMD64 \|\| X86 */
764
765	#define TEST_READ(a_pVar, a_Type, a_Fmt, a_Function, a_Val) \
766	do { a_pVar = a_Val; CHECKOP(a_Function(a_pVar), a_Val, a_Fmt, a_Type); CHECKVAL(a_pVar, a_Val, a_Fmt); } while (0)
767
768	DECLINLINE(void) tstASMAtomicReadU8Worker(uint8_t volatile *pu8)
769	{
770	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 0);
771	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 1);
772	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 2);
773	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 16);
774	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 32);
775	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 32);
776	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 127);
777	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 128);
778	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 169);
779	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 239);
780	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 254);
781	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicReadU8, 255);
782
783	int8_t volatile pi8 = (int8_t volatile )pu8;
784	TEST_READ(pi8, uint8_t, "%d", ASMAtomicReadS8, INT8_MAX);
785	TEST_READ(pi8, uint8_t, "%d", ASMAtomicReadS8, INT8_MIN);
786	TEST_READ(pi8, uint8_t, "%d", ASMAtomicReadS8, 42);
787	TEST_READ(pi8, uint8_t, "%d", ASMAtomicReadS8, -21);
788
789	bool volatile pf = (bool volatile )pu8;
790	TEST_READ(pf, bool, "%d", ASMAtomicReadBool, true);
791	TEST_READ(pf, bool, "%d", ASMAtomicReadBool, false);
792	}
793
794
795	DECLINLINE(void) tstASMAtomicUoReadU8Worker(uint8_t volatile *pu8)
796	{
797	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 0);
798	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 1);
799	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 2);
800	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 16);
801	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 32);
802	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 32);
803	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 127);
804	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 128);
805	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 169);
806	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 239);
807	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 254);
808	TEST_READ(pu8, uint8_t, "%#x", ASMAtomicUoReadU8, 255);
809
810	int8_t volatile pi8 = (int8_t volatile )pu8;
811	TEST_READ(pi8, uint8_t, "%d", ASMAtomicUoReadS8, INT8_MAX);
812	TEST_READ(pi8, uint8_t, "%d", ASMAtomicUoReadS8, INT8_MIN);
813	TEST_READ(pi8, uint8_t, "%d", ASMAtomicUoReadS8, 42);
814	TEST_READ(pi8, uint8_t, "%d", ASMAtomicUoReadS8, -21);
815
816	bool volatile pf = (bool volatile )pu8;
817	TEST_READ(pf, bool, "%d", ASMAtomicUoReadBool, true);
818	TEST_READ(pf, bool, "%d", ASMAtomicUoReadBool, false);
819	}
820
821
822	DECLINLINE(void) tstASMAtomicReadU16Worker(uint16_t volatile *pu16)
823	{
824	TEST_READ(pu16, uint16_t, "%#x", ASMAtomicReadU16, 0);
825	TEST_READ(pu16, uint16_t, "%#x", ASMAtomicReadU16, 19983);
826	TEST_READ(pu16, uint16_t, "%#x", ASMAtomicReadU16, INT16_MAX);
827	TEST_READ(pu16, uint16_t, "%#x", ASMAtomicReadU16, UINT16_MAX);
828
829	int16_t volatile pi16 = (int16_t volatile )pu16;
830	TEST_READ(pi16, uint16_t, "%d", ASMAtomicReadS16, INT16_MAX);
831	TEST_READ(pi16, uint16_t, "%d", ASMAtomicReadS16, INT16_MIN);
832	TEST_READ(pi16, uint16_t, "%d", ASMAtomicReadS16, 42);
833	TEST_READ(pi16, uint16_t, "%d", ASMAtomicReadS16, -21);
834	}
835
836
837	DECLINLINE(void) tstASMAtomicUoReadU16Worker(uint16_t volatile *pu16)
838	{
839	TEST_READ(pu16, uint16_t, "%#x", ASMAtomicUoReadU16, 0);
840	TEST_READ(pu16, uint16_t, "%#x", ASMAtomicUoReadU16, 19983);
841	TEST_READ(pu16, uint16_t, "%#x", ASMAtomicUoReadU16, INT16_MAX);
842	TEST_READ(pu16, uint16_t, "%#x", ASMAtomicUoReadU16, UINT16_MAX);
843
844	int16_t volatile pi16 = (int16_t volatile )pu16;
845	TEST_READ(pi16, uint16_t, "%d", ASMAtomicUoReadS16, INT16_MAX);
846	TEST_READ(pi16, uint16_t, "%d", ASMAtomicUoReadS16, INT16_MIN);
847	TEST_READ(pi16, uint16_t, "%d", ASMAtomicUoReadS16, 42);
848	TEST_READ(pi16, uint16_t, "%d", ASMAtomicUoReadS16, -21);
849	}
850
851
852	DECLINLINE(void) tstASMAtomicReadU32Worker(uint32_t volatile *pu32)
853	{
854	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicReadU32, 0);
855	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicReadU32, 19983);
856	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicReadU32, INT16_MAX);
857	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicReadU32, UINT16_MAX);
858	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicReadU32, _1M-1);
859	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicReadU32, _1M+1);
860	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicReadU32, _1G-1);
861	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicReadU32, _1G+1);
862	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicReadU32, INT32_MAX);
863	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicReadU32, UINT32_MAX);
864
865	int32_t volatile pi32 = (int32_t volatile )pu32;
866	TEST_READ(pi32, uint32_t, "%d", ASMAtomicReadS32, INT32_MAX);
867	TEST_READ(pi32, uint32_t, "%d", ASMAtomicReadS32, INT32_MIN);
868	TEST_READ(pi32, uint32_t, "%d", ASMAtomicReadS32, 42);
869	TEST_READ(pi32, uint32_t, "%d", ASMAtomicReadS32, -21);
870
871	#if ARCH_BITS == 32
872	size_t volatile pcb = (size_t volatile )pu32;
873	TEST_READ(pcb, size_t, "%#llz", ASMAtomicReadZ, 0);
874	TEST_READ(pcb, size_t, "%#llz", ASMAtomicReadZ, ~(size_t)2);
875	TEST_READ(pcb, size_t, "%#llz", ASMAtomicReadZ, ~(size_t)0 / 4);
876
877	void * volatile ppv = (void volatile *)pu32;
878	TEST_READ(ppv, void *, "%p", ASMAtomicReadPtr, NULL);
879	TEST_READ(ppv, void , "%p", ASMAtomicReadPtr, (void )~(uintptr_t)42);
880
881	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu32;
882	RTSEMEVENT hEvt = ASMAtomicReadPtrT(phEvt, RTSEMEVENT);
883	CHECKVAL(hEvt, (RTSEMEVENT)~(uintptr_t)42, "%p");
884
885	ASMAtomicReadHandle(phEvt, &hEvt);
886	CHECKVAL(hEvt, (RTSEMEVENT)~(uintptr_t)42, "%p");
887	#endif
888	}
889
890
891	DECLINLINE(void) tstASMAtomicUoReadU32Worker(uint32_t volatile *pu32)
892	{
893	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicUoReadU32, 0);
894	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicUoReadU32, 19983);
895	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicUoReadU32, INT16_MAX);
896	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicUoReadU32, UINT16_MAX);
897	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicUoReadU32, _1M-1);
898	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicUoReadU32, _1M+1);
899	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicUoReadU32, _1G-1);
900	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicUoReadU32, _1G+1);
901	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicUoReadU32, INT32_MAX);
902	TEST_READ(pu32, uint32_t, "%#x", ASMAtomicUoReadU32, UINT32_MAX);
903
904	int32_t volatile pi32 = (int32_t volatile )pu32;
905	TEST_READ(pi32, uint32_t, "%d", ASMAtomicUoReadS32, INT32_MAX);
906	TEST_READ(pi32, uint32_t, "%d", ASMAtomicUoReadS32, INT32_MIN);
907	TEST_READ(pi32, uint32_t, "%d", ASMAtomicUoReadS32, 42);
908	TEST_READ(pi32, uint32_t, "%d", ASMAtomicUoReadS32, -21);
909
910	#if ARCH_BITS == 32
911	size_t volatile pcb = (size_t volatile )pu32;
912	TEST_READ(pcb, size_t, "%#llz", ASMAtomicUoReadZ, 0);
913	TEST_READ(pcb, size_t, "%#llz", ASMAtomicUoReadZ, ~(size_t)2);
914	TEST_READ(pcb, size_t, "%#llz", ASMAtomicUoReadZ, ~(size_t)0 / 4);
915
916	void * volatile ppv = (void volatile *)pu32;
917	TEST_READ(ppv, void *, "%p", ASMAtomicUoReadPtr, NULL);
918	TEST_READ(ppv, void , "%p", ASMAtomicUoReadPtr, (void )~(uintptr_t)42);
919
920	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu32;
921	RTSEMEVENT hEvt = ASMAtomicUoReadPtrT(phEvt, RTSEMEVENT);
922	CHECKVAL(hEvt, (RTSEMEVENT)~(uintptr_t)42, "%p");
923
924	ASMAtomicUoReadHandle(phEvt, &hEvt);
925	CHECKVAL(hEvt, (RTSEMEVENT)~(uintptr_t)42, "%p");
926	#endif
927	}
928
929
930	DECLINLINE(void) tstASMAtomicReadU64Worker(uint64_t volatile *pu64)
931	{
932	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, 0);
933	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, 19983);
934	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, INT16_MAX);
935	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, UINT16_MAX);
936	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, _1M-1);
937	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, _1M+1);
938	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, _1G-1);
939	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, _1G+1);
940	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, INT32_MAX);
941	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, UINT32_MAX);
942	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, INT64_MAX);
943	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, UINT64_MAX);
944	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicReadU64, UINT64_C(0x450872549687134));
945
946	int64_t volatile pi64 = (int64_t volatile )pu64;
947	TEST_READ(pi64, uint64_t, "%d", ASMAtomicReadS64, INT64_MAX);
948	TEST_READ(pi64, uint64_t, "%d", ASMAtomicReadS64, INT64_MIN);
949	TEST_READ(pi64, uint64_t, "%d", ASMAtomicReadS64, 42);
950	TEST_READ(pi64, uint64_t, "%d", ASMAtomicReadS64, -21);
951
952	#if ARCH_BITS == 64
953	size_t volatile pcb = (size_t volatile )pu64;
954	TEST_READ(pcb, size_t, "%#llz", ASMAtomicReadZ, 0);
955	TEST_READ(pcb, size_t, "%#llz", ASMAtomicReadZ, ~(size_t)2);
956	TEST_READ(pcb, size_t, "%#llz", ASMAtomicReadZ, ~(size_t)0 / 4);
957
958	void * volatile ppv = (void volatile *)pu64;
959	TEST_READ(ppv, void *, "%p", ASMAtomicReadPtr, NULL);
960	TEST_READ(ppv, void , "%p", ASMAtomicReadPtr, (void )~(uintptr_t)42);
961
962	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu64;
963	RTSEMEVENT hEvt = ASMAtomicReadPtrT(phEvt, RTSEMEVENT);
964	CHECKVAL(hEvt, (RTSEMEVENT)~(uintptr_t)42, "%p");
965
966	ASMAtomicReadHandle(phEvt, &hEvt);
967	CHECKVAL(hEvt, (RTSEMEVENT)~(uintptr_t)42, "%p");
968	#endif
969	}
970
971
972	DECLINLINE(void) tstASMAtomicUoReadU64Worker(uint64_t volatile *pu64)
973	{
974	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, 0);
975	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, 19983);
976	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, INT16_MAX);
977	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, UINT16_MAX);
978	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, _1M-1);
979	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, _1M+1);
980	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, _1G-1);
981	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, _1G+1);
982	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, INT32_MAX);
983	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, UINT32_MAX);
984	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, INT64_MAX);
985	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, UINT64_MAX);
986	TEST_READ(pu64, uint64_t, "%#llx", ASMAtomicUoReadU64, UINT64_C(0x450872549687134));
987
988	int64_t volatile pi64 = (int64_t volatile )pu64;
989	TEST_READ(pi64, uint64_t, "%d", ASMAtomicUoReadS64, INT64_MAX);
990	TEST_READ(pi64, uint64_t, "%d", ASMAtomicUoReadS64, INT64_MIN);
991	TEST_READ(pi64, uint64_t, "%d", ASMAtomicUoReadS64, 42);
992	TEST_READ(pi64, uint64_t, "%d", ASMAtomicUoReadS64, -21);
993
994	#if ARCH_BITS == 64
995	size_t volatile pcb = (size_t volatile )pu64;
996	TEST_READ(pcb, size_t, "%#llz", ASMAtomicUoReadZ, 0);
997	TEST_READ(pcb, size_t, "%#llz", ASMAtomicUoReadZ, ~(size_t)2);
998	TEST_READ(pcb, size_t, "%#llz", ASMAtomicUoReadZ, ~(size_t)0 / 4);
999
1000	void * volatile ppv = (void volatile *)pu64;
1001	TEST_READ(ppv, void *, "%p", ASMAtomicUoReadPtr, NULL);
1002	TEST_READ(ppv, void , "%p", ASMAtomicUoReadPtr, (void )~(uintptr_t)42);
1003
1004	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu64;
1005	RTSEMEVENT hEvt = ASMAtomicUoReadPtrT(phEvt, RTSEMEVENT);
1006	CHECKVAL(hEvt, (RTSEMEVENT)~(uintptr_t)42, "%p");
1007
1008	ASMAtomicUoReadHandle(phEvt, &hEvt);
1009	CHECKVAL(hEvt, (RTSEMEVENT)~(uintptr_t)42, "%p");
1010	#endif
1011	}
1012
1013
1014	static void tstASMAtomicRead(void)
1015	{
1016	DO_SIMPLE_TEST(ASMAtomicReadU8, uint8_t);
1017	DO_SIMPLE_TEST(ASMAtomicUoReadU8, uint8_t);
1018
1019	DO_SIMPLE_TEST(ASMAtomicReadU16, uint16_t);
1020	DO_SIMPLE_TEST(ASMAtomicUoReadU16, uint16_t);
1021
1022	DO_SIMPLE_TEST(ASMAtomicReadU32, uint32_t);
1023	DO_SIMPLE_TEST(ASMAtomicUoReadU32, uint32_t);
1024
1025	DO_SIMPLE_TEST(ASMAtomicReadU64, uint64_t);
1026	DO_SIMPLE_TEST(ASMAtomicUoReadU64, uint64_t);
1027	}
1028
1029
1030	#define TEST_WRITE(a_pVar, a_Type, a_Fmt, a_Function, a_Val) \
1031	do { a_Function(a_pVar, a_Val); CHECKVAL(*a_pVar, a_Val, a_Fmt); } while (0)
1032
1033	DECLINLINE(void) tstASMAtomicWriteU8Worker(uint8_t volatile *pu8)
1034	{
1035	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 0);
1036	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 1);
1037	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 2);
1038	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 16);
1039	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 32);
1040	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 32);
1041	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 127);
1042	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 128);
1043	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 169);
1044	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 239);
1045	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 254);
1046	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicWriteU8, 255);
1047
1048	volatile int8_t pi8 = (volatile int8_t )pu8;
1049	TEST_WRITE(pi8, int8_t, "%d", ASMAtomicWriteS8, INT8_MIN);
1050	TEST_WRITE(pi8, int8_t, "%d", ASMAtomicWriteS8, INT8_MAX);
1051	TEST_WRITE(pi8, int8_t, "%d", ASMAtomicWriteS8, 42);
1052	TEST_WRITE(pi8, int8_t, "%d", ASMAtomicWriteS8, -41);
1053
1054	volatile bool pf = (volatile bool )pu8;
1055	TEST_WRITE(pf, bool, "%d", ASMAtomicWriteBool, true);
1056	TEST_WRITE(pf, bool, "%d", ASMAtomicWriteBool, false);
1057	}
1058
1059
1060	DECLINLINE(void) tstASMAtomicUoWriteU8Worker(uint8_t volatile *pu8)
1061	{
1062	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 0);
1063	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 1);
1064	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 2);
1065	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 16);
1066	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 32);
1067	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 32);
1068	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 127);
1069	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 128);
1070	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 169);
1071	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 239);
1072	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 254);
1073	TEST_WRITE(pu8, uint8_t, "%#x", ASMAtomicUoWriteU8, 255);
1074
1075	volatile int8_t pi8 = (volatile int8_t )pu8;
1076	TEST_WRITE(pi8, int8_t, "%d", ASMAtomicUoWriteS8, INT8_MIN);
1077	TEST_WRITE(pi8, int8_t, "%d", ASMAtomicUoWriteS8, INT8_MAX);
1078	TEST_WRITE(pi8, int8_t, "%d", ASMAtomicUoWriteS8, 42);
1079	TEST_WRITE(pi8, int8_t, "%d", ASMAtomicUoWriteS8, -41);
1080
1081	volatile bool pf = (volatile bool )pu8;
1082	TEST_WRITE(pf, bool, "%d", ASMAtomicUoWriteBool, true);
1083	TEST_WRITE(pf, bool, "%d", ASMAtomicUoWriteBool, false);
1084	}
1085
1086
1087	DECLINLINE(void) tstASMAtomicWriteU16Worker(uint16_t volatile *pu16)
1088	{
1089	TEST_WRITE(pu16, uint16_t, "%#x", ASMAtomicWriteU16, 0);
1090	TEST_WRITE(pu16, uint16_t, "%#x", ASMAtomicWriteU16, 19983);
1091	TEST_WRITE(pu16, uint16_t, "%#x", ASMAtomicWriteU16, INT16_MAX);
1092	TEST_WRITE(pu16, uint16_t, "%#x", ASMAtomicWriteU16, UINT16_MAX);
1093
1094	volatile int16_t pi16 = (volatile int16_t )pu16;
1095	TEST_WRITE(pi16, int16_t, "%d", ASMAtomicWriteS16, INT16_MIN);
1096	TEST_WRITE(pi16, int16_t, "%d", ASMAtomicWriteS16, INT16_MAX);
1097	TEST_WRITE(pi16, int16_t, "%d", ASMAtomicWriteS16, 42);
1098	TEST_WRITE(pi16, int16_t, "%d", ASMAtomicWriteS16, -41);
1099	}
1100
1101
1102	DECLINLINE(void) tstASMAtomicUoWriteU16Worker(uint16_t volatile *pu16)
1103	{
1104	TEST_WRITE(pu16, uint16_t, "%#x", ASMAtomicUoWriteU16, 0);
1105	TEST_WRITE(pu16, uint16_t, "%#x", ASMAtomicUoWriteU16, 19983);
1106	TEST_WRITE(pu16, uint16_t, "%#x", ASMAtomicUoWriteU16, INT16_MAX);
1107	TEST_WRITE(pu16, uint16_t, "%#x", ASMAtomicUoWriteU16, UINT16_MAX);
1108
1109	volatile int16_t pi16 = (volatile int16_t )pu16;
1110	TEST_WRITE(pi16, int16_t, "%d", ASMAtomicUoWriteS16, INT16_MIN);
1111	TEST_WRITE(pi16, int16_t, "%d", ASMAtomicUoWriteS16, INT16_MAX);
1112	TEST_WRITE(pi16, int16_t, "%d", ASMAtomicUoWriteS16, 42);
1113	TEST_WRITE(pi16, int16_t, "%d", ASMAtomicUoWriteS16, -41);
1114	}
1115
1116
1117	DECLINLINE(void) tstASMAtomicWriteU32Worker(uint32_t volatile *pu32)
1118	{
1119	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicWriteU32, 0);
1120	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicWriteU32, 19983);
1121	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicWriteU32, INT16_MAX);
1122	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicWriteU32, UINT16_MAX);
1123	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicWriteU32, _1M-1);
1124	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicWriteU32, _1M+1);
1125	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicWriteU32, _1G-1);
1126	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicWriteU32, _1G+1);
1127	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicWriteU32, INT32_MAX);
1128	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicWriteU32, UINT32_MAX);
1129
1130	volatile int32_t pi32 = (volatile int32_t )pu32;
1131	TEST_WRITE(pi32, int32_t, "%d", ASMAtomicWriteS32, INT32_MIN);
1132	TEST_WRITE(pi32, int32_t, "%d", ASMAtomicWriteS32, INT32_MAX);
1133	TEST_WRITE(pi32, int32_t, "%d", ASMAtomicWriteS32, 42);
1134	TEST_WRITE(pi32, int32_t, "%d", ASMAtomicWriteS32, -41);
1135
1136	#if ARCH_BITS == 32
1137	size_t volatile pcb = (size_t volatile )pu32;
1138	TEST_WRITE(pcb, size_t, "%#zx", ASMAtomicWriteZ, ~(size_t)42);
1139	TEST_WRITE(pcb, size_t, "%#zx", ASMAtomicWriteZ, 42);
1140
1141	void * volatile ppv = (void volatile *)pu32;
1142	TEST_WRITE(ppv, void *, "%#zx", ASMAtomicWritePtrVoid, NULL);
1143	TEST_WRITE(ppv, void , "%#zx", ASMAtomicWritePtrVoid, (void )~(uintptr_t)12938754);
1144
1145	ASMAtomicWriteNullPtr(ppv); CHECKVAL(*ppv, NULL, "%p");
1146	ASMAtomicWritePtr(ppv, (void )~(intptr_t)2322434); CHECKVAL(ppv, (void *)~(intptr_t)2322434, "%p");
1147
1148	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu32;
1149	ASMAtomicWriteHandle(phEvt, (RTSEMEVENT)(uintptr_t)99753456); CHECKVAL(*phEvt, (RTSEMEVENT)(uintptr_t)99753456, "%p");
1150	#endif
1151	}
1152
1153
1154	DECLINLINE(void) tstASMAtomicUoWriteU32Worker(uint32_t volatile *pu32)
1155	{
1156	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicUoWriteU32, 0);
1157	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicUoWriteU32, 19983);
1158	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicUoWriteU32, INT16_MAX);
1159	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicUoWriteU32, UINT16_MAX);
1160	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicUoWriteU32, _1M-1);
1161	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicUoWriteU32, _1M+1);
1162	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicUoWriteU32, _1G-1);
1163	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicUoWriteU32, _1G+1);
1164	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicUoWriteU32, INT32_MAX);
1165	TEST_WRITE(pu32, uint32_t, "%#x", ASMAtomicUoWriteU32, UINT32_MAX);
1166
1167	volatile int32_t pi32 = (volatile int32_t )pu32;
1168	TEST_WRITE(pi32, int32_t, "%d", ASMAtomicUoWriteS32, INT32_MIN);
1169	TEST_WRITE(pi32, int32_t, "%d", ASMAtomicUoWriteS32, INT32_MAX);
1170	TEST_WRITE(pi32, int32_t, "%d", ASMAtomicUoWriteS32, 42);
1171	TEST_WRITE(pi32, int32_t, "%d", ASMAtomicUoWriteS32, -41);
1172
1173	#if ARCH_BITS == 32
1174	size_t volatile pcb = (size_t volatile )pu32;
1175	TEST_WRITE(pcb, size_t, "%#zx", ASMAtomicUoWriteZ, ~(size_t)42);
1176	TEST_WRITE(pcb, size_t, "%#zx", ASMAtomicUoWriteZ, 42);
1177
1178	void * volatile ppv = (void volatile *)pu32;
1179	TEST_WRITE(ppv, void *, "%#zx", ASMAtomicUoWritePtrVoid, NULL);
1180	TEST_WRITE(ppv, void , "%#zx", ASMAtomicUoWritePtrVoid, (void )~(uintptr_t)12938754);
1181
1182	ASMAtomicUoWriteNullPtr(ppv); CHECKVAL(*ppv, NULL, "%p");
1183	ASMAtomicUoWritePtr(ppv, (void )~(intptr_t)2322434); CHECKVAL(ppv, (void *)~(intptr_t)2322434, "%p");
1184
1185	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu32;
1186	ASMAtomicUoWriteHandle(phEvt, (RTSEMEVENT)(uintptr_t)99753456); CHECKVAL(*phEvt, (RTSEMEVENT)(uintptr_t)99753456, "%p");
1187	#endif
1188	}
1189
1190
1191	DECLINLINE(void) tstASMAtomicWriteU64Worker(uint64_t volatile *pu64)
1192	{
1193	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, 0);
1194	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, 19983);
1195	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, INT16_MAX);
1196	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, UINT16_MAX);
1197	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, _1M-1);
1198	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, _1M+1);
1199	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, _1G-1);
1200	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, _1G+1);
1201	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, INT32_MAX);
1202	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, UINT32_MAX);
1203	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, INT64_MAX);
1204	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, UINT64_MAX);
1205	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicWriteU64, UINT64_C(0x450872549687134));
1206
1207	volatile int64_t pi64 = (volatile int64_t )pu64;
1208	TEST_WRITE(pi64, int64_t, "%d", ASMAtomicWriteS64, INT64_MIN);
1209	TEST_WRITE(pi64, int64_t, "%d", ASMAtomicWriteS64, INT64_MAX);
1210	TEST_WRITE(pi64, int64_t, "%d", ASMAtomicWriteS64, 42);
1211
1212	#if ARCH_BITS == 64
1213	size_t volatile pcb = (size_t volatile )pu64;
1214	TEST_WRITE(pcb, size_t, "%#zx", ASMAtomicWriteZ, ~(size_t)42);
1215	TEST_WRITE(pcb, size_t, "%#zx", ASMAtomicWriteZ, 42);
1216
1217	void * volatile ppv = (void volatile *)pu64;
1218	TEST_WRITE(ppv, void *, "%#zx", ASMAtomicWritePtrVoid, NULL);
1219	TEST_WRITE(ppv, void , "%#zx", ASMAtomicWritePtrVoid, (void )~(uintptr_t)12938754);
1220
1221	ASMAtomicWriteNullPtr(ppv); CHECKVAL(*ppv, NULL, "%p");
1222	ASMAtomicWritePtr(ppv, (void )~(intptr_t)2322434); CHECKVAL(ppv, (void *)~(intptr_t)2322434, "%p");
1223
1224	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu64;
1225	ASMAtomicWriteHandle(phEvt, (RTSEMEVENT)(uintptr_t)99753456); CHECKVAL(*phEvt, (RTSEMEVENT)(uintptr_t)99753456, "%p");
1226	#endif
1227	}
1228
1229
1230	DECLINLINE(void) tstASMAtomicUoWriteU64Worker(uint64_t volatile *pu64)
1231	{
1232	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, 0);
1233	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, 19983);
1234	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, INT16_MAX);
1235	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, UINT16_MAX);
1236	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, _1M-1);
1237	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, _1M+1);
1238	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, _1G-1);
1239	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, _1G+1);
1240	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, INT32_MAX);
1241	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, UINT32_MAX);
1242	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, INT64_MAX);
1243	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, UINT64_MAX);
1244	TEST_WRITE(pu64, uint64_t, "%#llx", ASMAtomicUoWriteU64, UINT64_C(0x450872549687134));
1245
1246	volatile int64_t pi64 = (volatile int64_t )pu64;
1247	TEST_WRITE(pi64, int64_t, "%d", ASMAtomicUoWriteS64, INT64_MIN);
1248	TEST_WRITE(pi64, int64_t, "%d", ASMAtomicUoWriteS64, INT64_MAX);
1249	TEST_WRITE(pi64, int64_t, "%d", ASMAtomicUoWriteS64, 42);
1250
1251	#if ARCH_BITS == 64
1252	size_t volatile pcb = (size_t volatile )pu64;
1253	TEST_WRITE(pcb, size_t, "%#zx", ASMAtomicUoWriteZ, ~(size_t)42);
1254	TEST_WRITE(pcb, size_t, "%#zx", ASMAtomicUoWriteZ, 42);
1255
1256	void * volatile ppv = (void volatile *)pu64;
1257	TEST_WRITE(ppv, void *, "%#zx", ASMAtomicUoWritePtrVoid, NULL);
1258	TEST_WRITE(ppv, void , "%#zx", ASMAtomicUoWritePtrVoid, (void )~(uintptr_t)12938754);
1259
1260	ASMAtomicUoWriteNullPtr(ppv); CHECKVAL(*ppv, NULL, "%p");
1261	ASMAtomicUoWritePtr(ppv, (void )~(intptr_t)2322434); CHECKVAL(ppv, (void *)~(intptr_t)2322434, "%p");
1262
1263	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu64;
1264	ASMAtomicUoWriteHandle(phEvt, (RTSEMEVENT)(uintptr_t)99753456); CHECKVAL(*phEvt, (RTSEMEVENT)(uintptr_t)99753456, "%p");
1265	#endif
1266	}
1267
1268	static void tstASMAtomicWrite(void)
1269	{
1270	DO_SIMPLE_TEST(ASMAtomicWriteU8, uint8_t);
1271	DO_SIMPLE_TEST(ASMAtomicUoWriteU8, uint8_t);
1272
1273	DO_SIMPLE_TEST(ASMAtomicWriteU16, uint16_t);
1274	DO_SIMPLE_TEST(ASMAtomicUoWriteU16, uint16_t);
1275
1276	DO_SIMPLE_TEST(ASMAtomicWriteU32, uint32_t);
1277	DO_SIMPLE_TEST(ASMAtomicUoWriteU32, uint32_t);
1278
1279	DO_SIMPLE_TEST(ASMAtomicWriteU64, uint64_t);
1280	DO_SIMPLE_TEST(ASMAtomicUoWriteU64, uint64_t);
1281	}
1282
1283
1284	DECLINLINE(void) tstASMAtomicXchgU8Worker(uint8_t volatile *pu8)
1285	{
1286	*pu8 = 0;
1287	CHECK_OP_AND_VAL(uint8_t, "%#x", pu8, ASMAtomicXchgU8(pu8, 1), 0, 1);
1288	CHECK_OP_AND_VAL(uint8_t, "%#x", pu8, ASMAtomicXchgU8(pu8, UINT8_C(0xff)), 1, UINT8_C(0xff));
1289	CHECK_OP_AND_VAL(uint8_t, "%#x", pu8, ASMAtomicXchgU8(pu8, UINT8_C(0x87)), UINT8_C(0xff), UINT8_C(0x87));
1290	CHECK_OP_AND_VAL(uint8_t, "%#x", pu8, ASMAtomicXchgU8(pu8, UINT8_C(0xfe)), UINT8_C(0x87), UINT8_C(0xfe));
1291
1292	int8_t volatile pi8 = (int8_t volatile )pu8;
1293	CHECK_OP_AND_VAL(int8_t, "%d", pi8, ASMAtomicXchgS8(pi8, INT8_C(-4)), INT8_C(-2), INT8_C(-4));
1294	CHECK_OP_AND_VAL(int8_t, "%d", pi8, ASMAtomicXchgS8(pi8, INT8_C(4)), INT8_C(-4), INT8_C(4));
1295	CHECK_OP_AND_VAL(int8_t, "%d", pi8, ASMAtomicXchgS8(pi8, INT8_MAX), INT8_C(4), INT8_MAX);
1296	CHECK_OP_AND_VAL(int8_t, "%d", pi8, ASMAtomicXchgS8(pi8, INT8_MIN), INT8_MAX, INT8_MIN);
1297	CHECK_OP_AND_VAL(int8_t, "%d", pi8, ASMAtomicXchgS8(pi8, 1), INT8_MIN, 1);
1298
1299	bool volatile pf = (bool volatile )pu8;
1300	CHECK_OP_AND_VAL(bool, "%d", pf, ASMAtomicXchgBool(pf, false), true, false);
1301	CHECK_OP_AND_VAL(bool, "%d", pf, ASMAtomicXchgBool(pf, false), false, false);
1302	CHECK_OP_AND_VAL(bool, "%d", pf, ASMAtomicXchgBool(pf, true), false, true);
1303	}
1304
1305
1306	DECLINLINE(void) tstASMAtomicXchgU16Worker(uint16_t volatile *pu16)
1307	{
1308	*pu16 = 0;
1309	CHECK_OP_AND_VAL(uint16_t, "%#x", pu16, ASMAtomicXchgU16(pu16, 1), 0, 1);
1310	CHECK_OP_AND_VAL(uint16_t, "%#x", pu16, ASMAtomicXchgU16(pu16, 0), 1, 0);
1311	CHECK_OP_AND_VAL(uint16_t, "%#x", pu16, ASMAtomicXchgU16(pu16, UINT16_MAX), 0, UINT16_MAX);
1312	CHECK_OP_AND_VAL(uint16_t, "%#x", pu16, ASMAtomicXchgU16(pu16, UINT16_C(0x7fff)), UINT16_MAX, UINT16_C(0x7fff));
1313	CHECK_OP_AND_VAL(uint16_t, "%#x", pu16, ASMAtomicXchgU16(pu16, UINT16_C(0x8765)), UINT16_C(0x7fff), UINT16_C(0x8765));
1314	CHECK_OP_AND_VAL(uint16_t, "%#x", pu16, ASMAtomicXchgU16(pu16, UINT16_C(0xfffe)), UINT16_C(0x8765), UINT16_C(0xfffe));
1315
1316	int16_t volatile pi16 = (int16_t volatile )pu16;
1317	CHECK_OP_AND_VAL(int16_t, "%d", pi16, ASMAtomicXchgS16(pi16, INT16_MIN), INT16_C(-2), INT16_MIN);
1318	CHECK_OP_AND_VAL(int16_t, "%d", pi16, ASMAtomicXchgS16(pi16, INT16_MAX), INT16_MIN, INT16_MAX);
1319	CHECK_OP_AND_VAL(int16_t, "%d", pi16, ASMAtomicXchgS16(pi16, -8), INT16_MAX, -8);
1320	CHECK_OP_AND_VAL(int16_t, "%d", pi16, ASMAtomicXchgS16(pi16, 8), -8, 8);
1321	}
1322
1323
1324	DECLINLINE(void) tstASMAtomicXchgU32Worker(uint32_t volatile *pu32)
1325	{
1326	*pu32 = 0;
1327	CHECK_OP_AND_VAL(uint32_t, "%#x", pu32, ASMAtomicXchgU32(pu32, 1), 0, 1);
1328	CHECK_OP_AND_VAL(uint32_t, "%#x", pu32, ASMAtomicXchgU32(pu32, 0), 1, 0);
1329	CHECK_OP_AND_VAL(uint32_t, "%#x", pu32, ASMAtomicXchgU32(pu32, UINT32_MAX), 0, UINT32_MAX);
1330	CHECK_OP_AND_VAL(uint32_t, "%#x", pu32, ASMAtomicXchgU32(pu32, UINT32_C(0x87654321)), UINT32_MAX, UINT32_C(0x87654321));
1331	CHECK_OP_AND_VAL(uint32_t, "%#x", pu32, ASMAtomicXchgU32(pu32, UINT32_C(0xfffffffe)), UINT32_C(0x87654321), UINT32_C(0xfffffffe));
1332
1333	int32_t volatile pi32 = (int32_t volatile )pu32;
1334	CHECK_OP_AND_VAL(int32_t, "%d", pi32, ASMAtomicXchgS32(pi32, INT32_MIN), INT32_C(-2), INT32_MIN);
1335	CHECK_OP_AND_VAL(int32_t, "%d", pi32, ASMAtomicXchgS32(pi32, INT32_MAX), INT32_MIN, INT32_MAX);
1336	CHECK_OP_AND_VAL(int32_t, "%d", pi32, ASMAtomicXchgS32(pi32, -16), INT32_MAX, -16);
1337	CHECK_OP_AND_VAL(int32_t, "%d", pi32, ASMAtomicXchgS32(pi32, 16), -16, 16);
1338
1339	#if ARCH_BITS == 32
1340	size_t volatile pcb = (size_t volatile )pu32;
1341	CHECK_OP_AND_VAL(size_t, "%#zx", pcb, ASMAtomicXchgZ(pcb, UINT32_C(0x9481239b)), 0x10, UINT32_C(0x9481239b));
1342	CHECK_OP_AND_VAL(size_t, "%#zx", pcb, ASMAtomicXchgZ(pcb, UINT32_C(0xcdef1234)), UINT32_C(0x9481239b), UINT32_C(0xcdef1234));
1343	#endif
1344
1345	#if R0_ARCH_BITS == 32
1346	RTR0PTR volatile pR0Ptr = (RTR0PTR volatile )pu32;
1347	CHECK_OP_AND_VAL(size_t, "%#llx", pcb, ASMAtomicXchgR0Ptr(pR0Ptr, UINT32_C(0x80341237)), UINT32_C(0xcdef1234), UINT32_C(0x80341237));
1348	#endif
1349	}
1350
1351
1352	DECLINLINE(void) tstASMAtomicXchgU64Worker(uint64_t volatile *pu64)
1353	{
1354	*pu64 = 0;
1355	CHECK_OP_AND_VAL(uint64_t, "%#llx", pu64, ASMAtomicXchgU64(pu64, 1), 0, 1);
1356	CHECK_OP_AND_VAL(uint64_t, "%#llx", pu64, ASMAtomicXchgU64(pu64, 0), 1, 0);
1357	CHECK_OP_AND_VAL(uint64_t, "%#llx", pu64, ASMAtomicXchgU64(pu64, UINT64_MAX), 0, UINT64_MAX);
1358	CHECK_OP_AND_VAL(uint64_t, "%#llx", pu64, ASMAtomicXchgU64(pu64, UINT64_C(0xfedcba0987654321)), UINT64_MAX, UINT64_C(0xfedcba0987654321));
1359	CHECK_OP_AND_VAL(uint64_t, "%#llx", pu64, ASMAtomicXchgU64(pu64, UINT64_C(0xfffffffffffffffe)), UINT64_C(0xfedcba0987654321), UINT64_C(0xfffffffffffffffe));
1360
1361	int64_t volatile pi64 = (int64_t volatile )pu64;
1362	CHECK_OP_AND_VAL(int64_t, "%lld", pi64, ASMAtomicXchgS64(pi64, INT64_MAX), -2, INT64_MAX);
1363	CHECK_OP_AND_VAL(int64_t, "%lld", pi64, ASMAtomicXchgS64(pi64, INT64_MIN), INT64_MAX, INT64_MIN);
1364	CHECK_OP_AND_VAL(int64_t, "%lld", pi64, ASMAtomicXchgS64(pi64, -32), INT64_MIN, -32);
1365	CHECK_OP_AND_VAL(int64_t, "%lld", pi64, ASMAtomicXchgS64(pi64, 32), -32, 32);
1366
1367	#if ARCH_BITS == 64
1368	size_t volatile pcb = (size_t volatile )pu64;
1369	CHECK_OP_AND_VAL(size_t, "%#zx", pcb, ASMAtomicXchgZ(pcb, UINT64_C(0x94812396759)), 0x20, UINT64_C(0x94812396759));
1370	CHECK_OP_AND_VAL(size_t, "%#zx", pcb, ASMAtomicXchgZ(pcb, UINT64_C(0xcdef1234abdf7896)), UINT64_C(0x94812396759), UINT64_C(0xcdef1234abdf7896));
1371	#endif
1372
1373	#if R0_ARCH_BITS == 64
1374	RTR0PTR volatile pR0Ptr = (RTR0PTR volatile )pu64;
1375	CHECK_OP_AND_VAL(size_t, "%#llx", pcb, ASMAtomicXchgR0Ptr(pR0Ptr, UINT64_C(0xfedc1234567890ab)), UINT64_C(0xcdef1234abdf7896), UINT64_C(0xfedc1234567890ab));
1376	#endif
1377	}
1378
1379
1380	DECLINLINE(void) tstASMAtomicXchgPtrWorker(void * volatile *ppv)
1381	{
1382	*ppv = NULL;
1383	CHECK_OP_AND_VAL(void , "%p", ppv, ASMAtomicXchgPtr(ppv, (void )(~(uintptr_t)0)), NULL, (void *)(~(uintptr_t)0));
1384	CHECK_OP_AND_VAL(void , "%p", ppv, ASMAtomicXchgPtr(ppv, (void )(~(uintptr_t)0x87654321)), (void )(~(uintptr_t)0), (void )(~(uintptr_t)0x87654321));
1385	CHECK_OP_AND_VAL(void , "%p", ppv, ASMAtomicXchgPtr(ppv, NULL), (void )(~(uintptr_t)0x87654321), NULL);
1386
1387	CHECK_OP_AND_VAL(void , "%p", ppv, ASMAtomicXchgR3Ptr(ppv, (void )ppv), NULL, (void *)ppv);
1388
1389	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )ppv;
1390	RTSEMEVENT hRet;
1391	ASMAtomicXchgHandle(phEvt, (RTSEMEVENT)(~(uintptr_t)12345), &hRet);
1392	CHECKVAL(hRet, (RTSEMEVENT)ppv, "%p");
1393	CHECKVAL(*phEvt, (RTSEMEVENT)(~(uintptr_t)12345), "%p");
1394	}
1395
1396
1397	static void tstASMAtomicXchg(void)
1398	{
1399	DO_SIMPLE_TEST(ASMAtomicXchgU8, uint8_t);
1400	DO_SIMPLE_TEST(ASMAtomicXchgU16, uint16_t);
1401	DO_SIMPLE_TEST(ASMAtomicXchgU32, uint32_t);
1402	DO_SIMPLE_TEST(ASMAtomicXchgU64, uint64_t);
1403	DO_SIMPLE_TEST(ASMAtomicXchgPtr, void *);
1404	}
1405
1406
1407	DECLINLINE(void) tstASMAtomicCmpXchgU8Worker(uint8_t volatile *pu8)
1408	{
1409	*pu8 = 0xff;
1410	CHECK_OP_AND_VAL_EX(bool, "%d", "%#x", pu8, ASMAtomicCmpXchgU8(pu8, 0, 0), false, 0xff);
1411	CHECK_OP_AND_VAL_EX(bool, "%d", "%#x", pu8, ASMAtomicCmpXchgU8(pu8, 0, 0xff), true, 0);
1412	CHECK_OP_AND_VAL_EX(bool, "%d", "%#x", pu8, ASMAtomicCmpXchgU8(pu8, 0x97, 0), true, 0x97);
1413	CHECK_OP_AND_VAL_EX(bool, "%d", "%#x", pu8, ASMAtomicCmpXchgU8(pu8, 0x97, 0), false, 0x97);
1414	CHECK_OP_AND_VAL_EX(bool, "%d", "%#x", pu8, ASMAtomicCmpXchgU8(pu8, 0x7f, 0x97), true, 0x7f);
1415
1416	int8_t volatile pi8 = (int8_t volatile )pu8;
1417	CHECK_OP_AND_VAL(bool, "%d", pi8, ASMAtomicCmpXchgS8(pi8, -2, 0x7f), true, -2);
1418	CHECK_OP_AND_VAL(bool, "%d", pi8, ASMAtomicCmpXchgS8(pi8, INT8_MAX, -2), true, INT8_MAX);
1419	CHECK_OP_AND_VAL(bool, "%d", pi8, ASMAtomicCmpXchgS8(pi8, INT8_MAX, INT8_MIN), false, INT8_MAX);
1420	CHECK_OP_AND_VAL(bool, "%d", pi8, ASMAtomicCmpXchgS8(pi8, INT8_MIN, INT8_MAX), true, INT8_MIN);
1421	CHECK_OP_AND_VAL(bool, "%d", pi8, ASMAtomicCmpXchgS8(pi8, 1, INT8_MIN), true, 1);
1422
1423	bool volatile pf = (bool volatile )pu8;
1424	CHECK_OP_AND_VAL(bool, "%d", pf, ASMAtomicCmpXchgBool(pf, true, true), true, true);
1425	CHECK_OP_AND_VAL(bool, "%d", pf, ASMAtomicCmpXchgBool(pf, false, true), true, false);
1426	CHECK_OP_AND_VAL(bool, "%d", pf, ASMAtomicCmpXchgBool(pf, false, true), false, false);
1427	CHECK_OP_AND_VAL(bool, "%d", pf, ASMAtomicCmpXchgBool(pf, false, false), true, false);
1428	}
1429
1430
1431	DECLINLINE(void) tstASMAtomicCmpXchgU32Worker(uint32_t volatile *pu32)
1432	{
1433	*pu32 = UINT32_C(0xffffffff);
1434	CHECK_OP_AND_VAL_EX(bool, "%d", "%#x", pu32, ASMAtomicCmpXchgU32(pu32, 0, 0), false, UINT32_C(0xffffffff));
1435	CHECK_OP_AND_VAL_EX(bool, "%d", "%#x", pu32, ASMAtomicCmpXchgU32(pu32, 0, UINT32_C(0xffffffff)), true, 0);
1436	CHECK_OP_AND_VAL_EX(bool, "%d", "%#x", pu32, ASMAtomicCmpXchgU32(pu32, UINT32_C(0x80088efd), UINT32_C(0x12345678)), false, 0);
1437	CHECK_OP_AND_VAL_EX(bool, "%d", "%#x", pu32, ASMAtomicCmpXchgU32(pu32, UINT32_C(0x80088efd), 0), true, UINT32_C(0x80088efd));
1438	CHECK_OP_AND_VAL_EX(bool, "%d", "%#x", pu32, ASMAtomicCmpXchgU32(pu32, UINT32_C(0xfffffffe), UINT32_C(0x80088efd)), true, UINT32_C(0xfffffffe));
1439
1440	int32_t volatile pi32 = (int32_t volatile )pu32;
1441	CHECK_OP_AND_VAL_EX(bool, "%d", "%d", pi32, ASMAtomicCmpXchgS32(pi32, INT32_MIN, 2), false, -2);
1442	CHECK_OP_AND_VAL_EX(bool, "%d", "%d", pi32, ASMAtomicCmpXchgS32(pi32, INT32_MIN, -2), true, INT32_MIN);
1443	CHECK_OP_AND_VAL_EX(bool, "%d", "%d", pi32, ASMAtomicCmpXchgS32(pi32, -19, -2), false, INT32_MIN);
1444	CHECK_OP_AND_VAL_EX(bool, "%d", "%d", pi32, ASMAtomicCmpXchgS32(pi32, -19, INT32_MIN), true, -19);
1445	CHECK_OP_AND_VAL_EX(bool, "%d", "%d", pi32, ASMAtomicCmpXchgS32(pi32, -19, INT32_MIN), false, -19);
1446	CHECK_OP_AND_VAL_EX(bool, "%d", "%d", pi32, ASMAtomicCmpXchgS32(pi32, 19, -19), true, 19);
1447	CHECK_OP_AND_VAL_EX(bool, "%d", "%d", pi32, ASMAtomicCmpXchgS32(pi32, INT32_MAX, -234), false, 19);
1448	CHECK_OP_AND_VAL_EX(bool, "%d", "%d", pi32, ASMAtomicCmpXchgS32(pi32, INT32_MAX, 19), true, INT32_MAX);
1449
1450	#if ARCH_BITS == 32
1451	void * volatile ppv = (void volatile *)pu32;
1452	CHECK_OP_AND_VAL_EX(bool, "%d", "%p", ppv, ASMAtomicCmpXchgPtrVoid(ppv, NULL, (void )(intptr_t)-29), false, (void )(intptr_t)29);
1453	CHECK_OP_AND_VAL_EX(bool, "%d", "%p", ppv, ASMAtomicCmpXchgPtrVoid(ppv, NULL, (void *)(intptr_t)29), true, NULL);
1454	CHECK_OP_AND_VAL_EX(bool, "%d", "%p", ppv, ASMAtomicCmpXchgPtrVoid(ppv, NULL, (void *)(intptr_t)29), false, NULL);
1455	CHECK_OP_AND_VAL_EX(bool, "%d", "%p", ppv, ASMAtomicCmpXchgPtrVoid(ppv, (void )~(uintptr_t)42, NULL), true, (void )~(uintptr_t)42);
1456
1457	bool fRc;
1458	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu32;
1459	ASMAtomicCmpXchgHandle(phEvt, (RTSEMEVENT)~(uintptr_t)0x12356389, (RTSEMEVENT)NULL, fRc);
1460	CHECKVAL(fRc, false, "%d");
1461	CHECKVAL(*phEvt, (RTSEMEVENT)~(uintptr_t)42, "%p");
1462
1463	ASMAtomicCmpXchgHandle(phEvt, (RTSEMEVENT)~(uintptr_t)0x12356389, (RTSEMEVENT)~(uintptr_t)42, fRc);
1464	CHECKVAL(fRc, true, "%d");
1465	CHECKVAL(*phEvt, (RTSEMEVENT)~(uintptr_t)0x12356389, "%p");
1466	#endif
1467	}
1468
1469
1470	DECLINLINE(void) tstASMAtomicCmpXchgU64Worker(uint64_t volatile *pu64)
1471	{
1472	*pu64 = UINT64_C(0xffffffffffffff);
1473	CHECK_OP_AND_VAL_EX(bool, "%d", "%#llx", pu64, ASMAtomicCmpXchgU64(pu64, 0, 0), false, UINT64_C(0xffffffffffffff));
1474	CHECK_OP_AND_VAL_EX(bool, "%d", "%#llx", pu64, ASMAtomicCmpXchgU64(pu64, 0, UINT64_C(0xffffffffffffff)), true, 0);
1475	CHECK_OP_AND_VAL_EX(bool, "%d", "%#llx", pu64, ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), 1), false, 0);
1476	CHECK_OP_AND_VAL_EX(bool, "%d", "%#llx", pu64, ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), 0), true, UINT64_C(0x80040008008efd));
1477	CHECK_OP_AND_VAL_EX(bool, "%d", "%#llx", pu64, ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), 0), false, UINT64_C(0x80040008008efd));
1478	CHECK_OP_AND_VAL_EX(bool, "%d", "%#llx", pu64, ASMAtomicCmpXchgU64(pu64, UINT64_C(0xfffffffffffffffd), UINT64_C(0x80040008008efd)), true, UINT64_C(0xfffffffffffffffd));
1479
1480	int64_t volatile pi64 = (int64_t volatile )pu64;
1481	CHECK_OP_AND_VAL_EX(bool, "%d", "%#lld", pi64, ASMAtomicCmpXchgS64(pi64, INT64_MAX, 0), false, -3);
1482	CHECK_OP_AND_VAL_EX(bool, "%d", "%#lld", pi64, ASMAtomicCmpXchgS64(pi64, INT64_MAX, -3), true, INT64_MAX);
1483	CHECK_OP_AND_VAL_EX(bool, "%d", "%#lld", pi64, ASMAtomicCmpXchgS64(pi64, INT64_MIN, INT64_MIN), false, INT64_MAX);
1484	CHECK_OP_AND_VAL_EX(bool, "%d", "%#lld", pi64, ASMAtomicCmpXchgS64(pi64, INT64_MIN, INT64_MAX), true, INT64_MIN);
1485	CHECK_OP_AND_VAL_EX(bool, "%d", "%#lld", pi64, ASMAtomicCmpXchgS64(pi64, -29, -29), false, INT64_MIN);
1486	CHECK_OP_AND_VAL_EX(bool, "%d", "%#lld", pi64, ASMAtomicCmpXchgS64(pi64, -29, INT64_MIN), true, -29);
1487	CHECK_OP_AND_VAL_EX(bool, "%d", "%#lld", pi64, ASMAtomicCmpXchgS64(pi64, -29, INT64_MIN), false, -29);
1488	CHECK_OP_AND_VAL_EX(bool, "%d", "%#lld", pi64, ASMAtomicCmpXchgS64(pi64, 29, -29), true, 29);
1489
1490	#if ARCH_BITS == 64
1491	void * volatile ppv = (void volatile *)pu64;
1492	CHECK_OP_AND_VAL_EX(bool, "%d", "%p", ppv, ASMAtomicCmpXchgPtrVoid(ppv, NULL, (void )(intptr_t)-29), false, (void )(intptr_t)29);
1493	CHECK_OP_AND_VAL_EX(bool, "%d", "%p", ppv, ASMAtomicCmpXchgPtrVoid(ppv, NULL, (void *)(intptr_t)29), true, NULL);
1494	CHECK_OP_AND_VAL_EX(bool, "%d", "%p", ppv, ASMAtomicCmpXchgPtrVoid(ppv, NULL, (void *)(intptr_t)29), false, NULL);
1495	CHECK_OP_AND_VAL_EX(bool, "%d", "%p", ppv, ASMAtomicCmpXchgPtrVoid(ppv, (void )~(uintptr_t)42, NULL), true, (void )~(uintptr_t)42);
1496
1497	bool fRc;
1498	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu64;
1499	ASMAtomicCmpXchgHandle(phEvt, (RTSEMEVENT)~(uintptr_t)0x12356389, (RTSEMEVENT)NULL, fRc);
1500	CHECKVAL(fRc, false, "%d");
1501	CHECKVAL(*phEvt, (RTSEMEVENT)~(uintptr_t)42, "%p");
1502
1503	ASMAtomicCmpXchgHandle(phEvt, (RTSEMEVENT)~(uintptr_t)0x12356389, (RTSEMEVENT)~(uintptr_t)42, fRc);
1504	CHECKVAL(fRc, true, "%d");
1505	CHECKVAL(*phEvt, (RTSEMEVENT)~(uintptr_t)0x12356389, "%p");
1506	#endif
1507	}
1508
1509
1510	#ifdef RTASM_HAVE_CMP_WRITE_U128
1511	DECLINLINE(void) tstASMAtomicCmpWriteU128Worker(RTUINT128U volatile *pu128)
1512	{
1513	pu128->s.Lo = UINT64_C(0xffffffffffffff);
1514	pu128->s.Hi = UINT64_C(0xffffffffffffff);
1515
1516	RTUINT128U u128A, u128B;
1517	CHECK_OP_AND_VAL_128_C(bool, "%d", pu128, ASMAtomicCmpWriteU128U(pu128,
1518	u128A = RTUINT128_INIT_C(0, 0),
1519	u128B = RTUINT128_INIT_C(0, 0)),
1520	false, 0xffffffffffffff, 0xffffffffffffff);
1521	CHECK_OP_AND_VAL_128_C(bool, "%d", pu128, ASMAtomicCmpWriteU128U(pu128,
1522	u128A = RTUINT128_INIT_C(0, 0),
1523	u128B = RTUINT128_INIT_C(0xffffffffffffff, 0xffffffffffffff)),
1524	true, 0, 0);
1525
1526	CHECK_OP_AND_VAL_128_C(bool, "%d", pu128, ASMAtomicCmpWriteU128U(pu128,
1527	u128A = RTUINT128_INIT_C(0x80040008008efd, 0x40080004004def),
1528	u128B = RTUINT128_INIT_C(0, 1)),
1529	false, 0, 0);
1530	CHECK_OP_AND_VAL_128_C(bool, "%d", pu128, ASMAtomicCmpWriteU128U(pu128,
1531	u128A = RTUINT128_INIT_C(0x80040008008efd, 0x40080004004def),
1532	u128B = RTUINT128_INIT_C(1, 0)),
1533	false, 0, 0);
1534	CHECK_OP_AND_VAL_128_C(bool, "%d", pu128, ASMAtomicCmpWriteU128U(pu128,
1535	u128A = RTUINT128_INIT_C(0x80040008008efd, 0x40080004004def),
1536	u128B = RTUINT128_INIT_C(0, 0)),
1537	true, 0x80040008008efd, 0x40080004004def);
1538
1539	CHECK_OP_AND_VAL_128_C(bool, "%d", pu128, ASMAtomicCmpWriteU128U(pu128,
1540	u128A = RTUINT128_INIT_C(0xfff40ff8f08ef3, 0x4ee8ee04cc4de4),
1541	u128B = RTUINT128_INIT_C(0x80040008008efd, 0)),
1542	false, 0x80040008008efd, 0x40080004004def);
1543	CHECK_OP_AND_VAL_128_C(bool, "%d", pu128, ASMAtomicCmpWriteU128U(pu128,
1544	u128A = RTUINT128_INIT_C(0xfff40ff8f08ef3, 0x4ee8ee04cc4de4),
1545	u128B = RTUINT128_INIT_C(0, 0x40080004004def)),
1546	false, 0x80040008008efd, 0x40080004004def);
1547	CHECK_OP_AND_VAL_128_C(bool, "%d", pu128, ASMAtomicCmpWriteU128U(pu128,
1548	u128A = RTUINT128_INIT_C(0xfff40ff8f08ef3, 0x4ee8ee04cc4de4),
1549	u128B = RTUINT128_INIT_C(0x80040008008efd, 0x40080004004def)),
1550	true, 0xfff40ff8f08ef3, 0x4ee8ee04cc4de4);
1551	}
1552	#endif /* RTASM_HAVE_CMP_WRITE_U128 */
1553
1554
1555	static void tstASMAtomicCmpXchg(void)
1556	{
1557	DO_SIMPLE_TEST(ASMAtomicCmpXchgU8, uint8_t);
1558	DO_SIMPLE_TEST(ASMAtomicCmpXchgU32, uint32_t);
1559	DO_SIMPLE_TEST(ASMAtomicCmpXchgU64, uint64_t);
1560	#ifdef RTASM_HAVE_CMP_WRITE_U128
1561	# ifdef RT_ARCH_AMD64
1562	if (ASMCpuId_ECX(1) & X86_CPUID_FEATURE_ECX_CX16)
1563	# endif
1564	{
1565	RTTestISub("ASMAtomicCmpWriteU128U");
1566	DO_SIMPLE_TEST_NO_SUB_NO_STACK(tstASMAtomicCmpWriteU128Worker, RTUINT128U);
1567	}
1568	#endif
1569	}
1570
1571
1572	DECLINLINE(void) tstASMAtomicCmpXchgExU8Worker(uint8_t volatile *pu8)
1573	{
1574	*pu8 = UINT8_C(0xff);
1575	uint8_t u8Old = UINT8_C(0x11);
1576	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu8, u8Old, ASMAtomicCmpXchgExU8(pu8, 0, 0, &u8Old), false, UINT8_C(0xff), UINT8_C(0xff));
1577	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu8, u8Old, ASMAtomicCmpXchgExU8(pu8, 0, UINT8_C(0xff), &u8Old), true, 0, UINT8_C(0xff));
1578	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu8, u8Old, ASMAtomicCmpXchgExU8(pu8, 0, UINT8_C(0xff), &u8Old), false, 0, UINT8_C(0x00));
1579	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu8, u8Old, ASMAtomicCmpXchgExU8(pu8, UINT8_C(0xfd), 0, &u8Old), true, UINT8_C(0xfd), 0);
1580	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu8, u8Old, ASMAtomicCmpXchgExU8(pu8, UINT8_C(0xfd), 0, &u8Old), false, UINT8_C(0xfd), UINT8_C(0xfd));
1581	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu8, u8Old, ASMAtomicCmpXchgExU8(pu8, UINT8_C(0xe0), UINT8_C(0xfd), &u8Old), true, UINT8_C(0xe0), UINT8_C(0xfd));
1582
1583	int8_t volatile pi8 = (int8_t volatile )pu8;
1584	int8_t i8Old = 0;
1585	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi8, i8Old, ASMAtomicCmpXchgExS8(pi8, 32, 32, &i8Old), false, -32, -32);
1586	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi8, i8Old, ASMAtomicCmpXchgExS8(pi8, 32, -32, &i8Old), true, 32, -32);
1587	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi8, i8Old, ASMAtomicCmpXchgExS8(pi8, INT8_MIN, 32, &i8Old), true, INT8_MIN, 32);
1588	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi8, i8Old, ASMAtomicCmpXchgExS8(pi8, INT8_MIN, 32, &i8Old), false, INT8_MIN, INT8_MIN);
1589	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi8, i8Old, ASMAtomicCmpXchgExS8(pi8, INT8_MAX, INT8_MAX, &i8Old), false, INT8_MIN, INT8_MIN);
1590	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi8, i8Old, ASMAtomicCmpXchgExS8(pi8, INT8_MAX, INT8_MIN, &i8Old), true, INT8_MAX, INT8_MIN);
1591	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi8, i8Old, ASMAtomicCmpXchgExS8(pi8, 42, INT8_MAX, &i8Old), true, 42, INT8_MAX);
1592	}
1593
1594
1595	DECLINLINE(void) tstASMAtomicCmpXchgExU16Worker(uint16_t volatile *pu16)
1596	{
1597	*pu16 = UINT16_C(0xffff);
1598	uint16_t u16Old = UINT16_C(0x5111);
1599	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu16, u16Old, ASMAtomicCmpXchgExU16(pu16, 0, 0, &u16Old), false, UINT16_C(0xffff), UINT16_C(0xffff));
1600	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu16, u16Old, ASMAtomicCmpXchgExU16(pu16, 0, UINT16_C(0xffff), &u16Old), true, 0, UINT16_C(0xffff));
1601	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu16, u16Old, ASMAtomicCmpXchgExU16(pu16, 0, UINT16_C(0xffff), &u16Old), false, 0, UINT16_C(0x0000));
1602	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu16, u16Old, ASMAtomicCmpXchgExU16(pu16, UINT16_C(0x8efd), 0, &u16Old), true, UINT16_C(0x8efd), 0);
1603	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu16, u16Old, ASMAtomicCmpXchgExU16(pu16, UINT16_C(0x8efd), 0, &u16Old), false, UINT16_C(0x8efd), UINT16_C(0x8efd));
1604	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu16, u16Old, ASMAtomicCmpXchgExU16(pu16, UINT16_C(0xffe0), UINT16_C(0x8efd), &u16Old), true, UINT16_C(0xffe0), UINT16_C(0x8efd));
1605
1606	int16_t volatile pi16 = (int16_t volatile )pu16;
1607	int16_t i16Old = 0;
1608	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi16, i16Old, ASMAtomicCmpXchgExS16(pi16, 32, 32, &i16Old), false, -32, -32);
1609	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi16, i16Old, ASMAtomicCmpXchgExS16(pi16, 32, -32, &i16Old), true, 32, -32);
1610	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi16, i16Old, ASMAtomicCmpXchgExS16(pi16, INT16_MIN, 32, &i16Old), true, INT16_MIN, 32);
1611	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi16, i16Old, ASMAtomicCmpXchgExS16(pi16, INT16_MIN, 32, &i16Old), false, INT16_MIN, INT16_MIN);
1612	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi16, i16Old, ASMAtomicCmpXchgExS16(pi16, INT16_MAX, INT16_MAX, &i16Old), false, INT16_MIN, INT16_MIN);
1613	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi16, i16Old, ASMAtomicCmpXchgExS16(pi16, INT16_MAX, INT16_MIN, &i16Old), true, INT16_MAX, INT16_MIN);
1614	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi16, i16Old, ASMAtomicCmpXchgExS16(pi16, 42, INT16_MAX, &i16Old), true, 42, INT16_MAX);
1615	}
1616
1617
1618	DECLINLINE(void) tstASMAtomicCmpXchgExU32Worker(uint32_t volatile *pu32)
1619	{
1620	*pu32 = UINT32_C(0xffffffff);
1621	uint32_t u32Old = UINT32_C(0x80005111);
1622	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu32, u32Old, ASMAtomicCmpXchgExU32(pu32, 0, 0, &u32Old), false, UINT32_C(0xffffffff), UINT32_C(0xffffffff));
1623	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu32, u32Old, ASMAtomicCmpXchgExU32(pu32, 0, UINT32_C(0xffffffff), &u32Old), true, 0, UINT32_C(0xffffffff));
1624	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu32, u32Old, ASMAtomicCmpXchgExU32(pu32, 0, UINT32_C(0xffffffff), &u32Old), false, 0, UINT32_C(0x00000000));
1625	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu32, u32Old, ASMAtomicCmpXchgExU32(pu32, UINT32_C(0x80088efd), 0, &u32Old), true, UINT32_C(0x80088efd), 0);
1626	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu32, u32Old, ASMAtomicCmpXchgExU32(pu32, UINT32_C(0x80088efd), 0, &u32Old), false, UINT32_C(0x80088efd), UINT32_C(0x80088efd));
1627	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#x", pu32, u32Old, ASMAtomicCmpXchgExU32(pu32, UINT32_C(0xffffffe0), UINT32_C(0x80088efd), &u32Old), true, UINT32_C(0xffffffe0), UINT32_C(0x80088efd));
1628
1629	int32_t volatile pi32 = (int32_t volatile )pu32;
1630	int32_t i32Old = 0;
1631	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi32, i32Old, ASMAtomicCmpXchgExS32(pi32, 32, 32, &i32Old), false, -32, -32);
1632	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi32, i32Old, ASMAtomicCmpXchgExS32(pi32, 32, -32, &i32Old), true, 32, -32);
1633	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi32, i32Old, ASMAtomicCmpXchgExS32(pi32, INT32_MIN, 32, &i32Old), true, INT32_MIN, 32);
1634	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi32, i32Old, ASMAtomicCmpXchgExS32(pi32, INT32_MIN, 32, &i32Old), false, INT32_MIN, INT32_MIN);
1635	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi32, i32Old, ASMAtomicCmpXchgExS32(pi32, INT32_MAX, INT32_MAX, &i32Old), false, INT32_MIN, INT32_MIN);
1636	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi32, i32Old, ASMAtomicCmpXchgExS32(pi32, INT32_MAX, INT32_MIN, &i32Old), true, INT32_MAX, INT32_MIN);
1637	CHECK_OP_AND_VAL_EX2(bool, "%d", "%d", pi32, i32Old, ASMAtomicCmpXchgExS32(pi32, 42, INT32_MAX, &i32Old), true, 42, INT32_MAX);
1638
1639	#if ARCH_BITS == 32
1640	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu32;
1641	RTSEMEVENT hEvtOld = (RTSEMEVENT)~(uintptr_t)31;
1642	bool fRc = true;
1643	ASMAtomicCmpXchgExHandle(phEvt, (RTSEMEVENT)~(uintptr_t)0x12380964, (RTSEMEVENT)~(uintptr_t)0, fRc, &hEvtOld);
1644	CHECKVAL(fRc, false, "%d");
1645	CHECKVAL(*phEvt, (RTSEMEVENT)(uintptr_t)42, "%p");
1646	CHECKVAL(hEvtOld, (RTSEMEVENT)(uintptr_t)42, "%p");
1647
1648	ASMAtomicCmpXchgExHandle(phEvt, (RTSEMEVENT)~(uintptr_t)0x12380964, (RTSEMEVENT)(uintptr_t)42, fRc, &hEvtOld);
1649	CHECKVAL(fRc, true, "%d");
1650	CHECKVAL(*phEvt, (RTSEMEVENT)~(uintptr_t)0x12380964, "%p");
1651	CHECKVAL(hEvtOld, (RTSEMEVENT)(uintptr_t)42, "%p");
1652	#endif
1653	}
1654
1655
1656	DECLINLINE(void) tstASMAtomicCmpXchgExU64Worker(uint64_t volatile *pu64)
1657	{
1658	*pu64 = UINT64_C(0xffffffffffffffff);
1659	uint64_t u64Old = UINT64_C(0x8000000051111111);
1660	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#llx", pu64, u64Old, ASMAtomicCmpXchgExU64(pu64, 0, 0, &u64Old), false, UINT64_C(0xffffffffffffffff), UINT64_C(0xffffffffffffffff));
1661	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#llx", pu64, u64Old, ASMAtomicCmpXchgExU64(pu64, 0, UINT64_C(0xffffffffffffffff), &u64Old), true, 0, UINT64_C(0xffffffffffffffff));
1662	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#llx", pu64, u64Old, ASMAtomicCmpXchgExU64(pu64, UINT64_C(0x0080040008008efd), 0x342, &u64Old), false, 0, 0);
1663	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#llx", pu64, u64Old, ASMAtomicCmpXchgExU64(pu64, UINT64_C(0x0080040008008efd), 0, &u64Old), true, UINT64_C(0x0080040008008efd), 0);
1664	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#llx", pu64, u64Old, ASMAtomicCmpXchgExU64(pu64, UINT64_C(0xffffffffffffffc0), UINT64_C(0x0080040008008efd), &u64Old), true, UINT64_C(0xffffffffffffffc0), UINT64_C(0x0080040008008efd));
1665
1666	int64_t volatile pi64 = (int64_t volatile )pu64;
1667	int64_t i64Old = -3;
1668	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#lld", pi64, i64Old, ASMAtomicCmpXchgExS64(pi64, 64, 64, &i64Old), false, -64, -64);
1669	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#lld", pi64, i64Old, ASMAtomicCmpXchgExS64(pi64, 64, -64, &i64Old), true, 64, -64);
1670	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#lld", pi64, i64Old, ASMAtomicCmpXchgExS64(pi64, 64, -64, &i64Old), false, 64, 64);
1671	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#lld", pi64, i64Old, ASMAtomicCmpXchgExS64(pi64, INT64_MIN, -64, &i64Old), false, 64, 64);
1672	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#lld", pi64, i64Old, ASMAtomicCmpXchgExS64(pi64, INT64_MIN, 64, &i64Old), true, INT64_MIN, 64);
1673	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#lld", pi64, i64Old, ASMAtomicCmpXchgExS64(pi64, INT64_MAX, INT64_MIN, &i64Old), true, INT64_MAX, INT64_MIN);
1674	CHECK_OP_AND_VAL_EX2(bool, "%d", "%#lld", pi64, i64Old, ASMAtomicCmpXchgExS64(pi64, 42, INT64_MAX, &i64Old), true, 42, INT64_MAX);
1675
1676	#if ARCH_BITS == 64
1677	RTSEMEVENT volatile phEvt = (RTSEMEVENT volatile )pu64;
1678	RTSEMEVENT hEvtOld = (RTSEMEVENT)~(uintptr_t)31;
1679	bool fRc = true;
1680	ASMAtomicCmpXchgExHandle(phEvt, (RTSEMEVENT)~(uintptr_t)0x12380964, (RTSEMEVENT)~(uintptr_t)0, fRc, &hEvtOld);
1681	CHECKVAL(fRc, false, "%d");
1682	CHECKVAL(*phEvt, (RTSEMEVENT)(uintptr_t)42, "%p");
1683	CHECKVAL(hEvtOld, (RTSEMEVENT)(uintptr_t)42, "%p");
1684
1685	ASMAtomicCmpXchgExHandle(phEvt, (RTSEMEVENT)~(uintptr_t)0x12380964, (RTSEMEVENT)(uintptr_t)42, fRc, &hEvtOld);
1686	CHECKVAL(fRc, true, "%d");
1687	CHECKVAL(*phEvt, (RTSEMEVENT)~(uintptr_t)0x12380964, "%p");
1688	CHECKVAL(hEvtOld, (RTSEMEVENT)(uintptr_t)42, "%p");
1689
1690	void * volatile ppv = (void volatile *)pu64;
1691	void *pvOld;
1692	CHECK_OP_AND_VAL_EX2(bool, "%d", "%p", ppv, pvOld, ASMAtomicCmpXchgExPtrVoid(ppv, (void )(intptr_t)12345678, NULL, &pvOld), false, (void )~(uintptr_t)0x12380964, (void *)~(uintptr_t)0x12380964);
1693	CHECK_OP_AND_VAL_EX2(bool, "%d", "%p", ppv, pvOld, ASMAtomicCmpXchgExPtrVoid(ppv, (void )(intptr_t)12345678, (void )~(uintptr_t)0x12380964, &pvOld), true, (void )(intptr_t)12345678, (void )~(uintptr_t)0x12380964);
1694
1695	CHECK_OP_AND_VAL_EX2(bool, "%d", "%p", ppv, pvOld, ASMAtomicCmpXchgExPtr(ppv, (void )~(uintptr_t)99, (void )~(uintptr_t)99, &pvOld), false, (void )(intptr_t)12345678, (void )(intptr_t)12345678);
1696	CHECK_OP_AND_VAL_EX2(bool, "%d", "%p", ppv, pvOld, ASMAtomicCmpXchgExPtr(ppv, (void )~(uintptr_t)99, (void )(intptr_t)12345678, &pvOld), true, (void )~(intptr_t)99, (void )(intptr_t)12345678);
1697	#endif
1698	}
1699
1700
1701	static void tstASMAtomicCmpXchgEx(void)
1702	{
1703	DO_SIMPLE_TEST(ASMAtomicCmpXchgExU8, uint8_t);
1704	DO_SIMPLE_TEST(ASMAtomicCmpXchgExU16, uint16_t);
1705	DO_SIMPLE_TEST(ASMAtomicCmpXchgExU32, uint32_t);
1706	DO_SIMPLE_TEST(ASMAtomicCmpXchgExU64, uint64_t);
1707	}
1708
1709
1710	#define TEST_RET_OLD(a_Type, a_Fmt, a_pVar, a_Function, a_uVal, a_VarExpect) do { \
1711	a_Type const uOldExpect = *(a_pVar); \
1712	a_Type uOldRet = a_Function(a_pVar, a_uVal); \
1713	if (RT_LIKELY( uOldRet == (uOldExpect) && *(a_pVar) == (a_VarExpect) )) { } \
1714	else RTTestFailed(g_hTest, "%s, %d: FAILURE: %s(%s," a_Fmt ") -> " a_Fmt ", expected " a_Fmt "; %s=" a_Fmt ", expected " a_Fmt "\n", \
1715	__FUNCTION__, __LINE__, #a_Function, #a_pVar, a_uVal, uOldRet, uOldExpect, #a_pVar, *(a_pVar), (a_VarExpect)); \
1716	} while (0)
1717
1718
1719	DECLINLINE(void) tstASMAtomicAddU32Worker(uint32_t *pu32)
1720	{
1721	*pu32 = 10;
1722	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAddU32, 1, 11);
1723	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAddU32, UINT32_C(0xfffffffe), 9);
1724	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAddU32, UINT32_C(0xfffffff7), 0);
1725	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAddU32, UINT32_C(0x7fffffff), UINT32_C(0x7fffffff));
1726	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAddU32, 1, UINT32_C(0x80000000));
1727	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAddU32, 1, UINT32_C(0x80000001));
1728	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAddU32, UINT32_C(0x7fffffff), 0);
1729	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAddU32, 0, 0);
1730
1731	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicSubU32, 0, 0);
1732	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicSubU32, 32, UINT32_C(0xffffffe0));
1733	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicSubU32, UINT32_C(0x7fffffff), UINT32_C(0x7fffffe1));
1734	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicSubU32, UINT32_C(0x7fffffde), UINT32_C(0x00000003));
1735	}
1736
1737
1738	DECLINLINE(void) tstASMAtomicAddS32Worker(int32_t *pi32)
1739	{
1740	*pi32 = 10;
1741	TEST_RET_OLD(int32_t, "%d", pi32, ASMAtomicAddS32, 1, 11);
1742	TEST_RET_OLD(int32_t, "%d", pi32, ASMAtomicAddS32, -2, 9);
1743	TEST_RET_OLD(int32_t, "%d", pi32, ASMAtomicAddS32, -9, 0);
1744	TEST_RET_OLD(int32_t, "%d", pi32, ASMAtomicAddS32, -0x7fffffff, -0x7fffffff);
1745	TEST_RET_OLD(int32_t, "%d", pi32, ASMAtomicAddS32, 0, -0x7fffffff);
1746	TEST_RET_OLD(int32_t, "%d", pi32, ASMAtomicAddS32, 0x7fffffff, 0);
1747	TEST_RET_OLD(int32_t, "%d", pi32, ASMAtomicAddS32, 0, 0);
1748
1749	TEST_RET_OLD(int32_t, "%d", pi32, ASMAtomicSubS32, 0, 0);
1750	TEST_RET_OLD(int32_t, "%d", pi32, ASMAtomicSubS32, 1, -1);
1751	TEST_RET_OLD(int32_t, "%d", pi32, ASMAtomicSubS32, INT32_MIN, INT32_MAX);
1752	}
1753
1754
1755	DECLINLINE(void) tstASMAtomicAddU64Worker(uint64_t volatile *pu64)
1756	{
1757	*pu64 = 10;
1758	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicAddU64, 1, 11);
1759	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicAddU64, UINT64_C(0xfffffffffffffffe), UINT64_C(0x0000000000000009));
1760	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicAddU64, UINT64_C(0xfffffffffffffff7), UINT64_C(0x0000000000000000));
1761	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicAddU64, UINT64_C(0x7ffffffffffffff0), UINT64_C(0x7ffffffffffffff0));
1762	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicAddU64, UINT64_C(0x7ffffffffffffff0), UINT64_C(0xffffffffffffffe0));
1763	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicAddU64, UINT64_C(0x0000000000000000), UINT64_C(0xffffffffffffffe0));
1764	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicAddU64, UINT64_C(0x000000000000001f), UINT64_C(0xffffffffffffffff));
1765	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicAddU64, UINT64_C(0x0000000000000001), UINT64_C(0x0000000000000000));
1766
1767	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicSubU64, UINT64_C(0x0000000000000000), UINT64_C(0x0000000000000000));
1768	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicSubU64, UINT64_C(0x0000000000000020), UINT64_C(0xffffffffffffffe0));
1769	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicSubU64, UINT64_C(0x7fffffffffffffff), UINT64_C(0x7fffffffffffffe1));
1770	TEST_RET_OLD(uint64_t, "%llx", pu64, ASMAtomicSubU64, UINT64_C(0x7fffffffffffffdd), UINT64_C(0x0000000000000004));
1771	}
1772
1773
1774	DECLINLINE(void) tstASMAtomicAddS64Worker(int64_t volatile *pi64)
1775	{
1776	*pi64 = 10;
1777	TEST_RET_OLD(int64_t, "%lld", pi64, ASMAtomicAddS64, 1, 11);
1778	TEST_RET_OLD(int64_t, "%lld", pi64, ASMAtomicAddS64, -2, 9);
1779	TEST_RET_OLD(int64_t, "%lld", pi64, ASMAtomicAddS64, -9, 0);
1780	TEST_RET_OLD(int64_t, "%lld", pi64, ASMAtomicAddS64, -INT64_MAX, -INT64_MAX);
1781	TEST_RET_OLD(int64_t, "%lld", pi64, ASMAtomicAddS64, 0, -INT64_MAX);
1782	TEST_RET_OLD(int64_t, "%lld", pi64, ASMAtomicAddS64, -1, INT64_MIN);
1783	TEST_RET_OLD(int64_t, "%lld", pi64, ASMAtomicAddS64, INT64_MAX, -1);
1784	TEST_RET_OLD(int64_t, "%lld", pi64, ASMAtomicAddS64, 1, 0);
1785	TEST_RET_OLD(int64_t, "%lld", pi64, ASMAtomicAddS64, 0, 0);
1786
1787	TEST_RET_OLD(int64_t, "%d", pi64, ASMAtomicSubS64, 0, 0);
1788	TEST_RET_OLD(int64_t, "%d", pi64, ASMAtomicSubS64, 1, -1);
1789	TEST_RET_OLD(int64_t, "%d", pi64, ASMAtomicSubS64, INT64_MIN, INT64_MAX);
1790	}
1791
1792
1793
1794	DECLINLINE(void) tstASMAtomicAddZWorker(size_t volatile *pcb)
1795	{
1796	*pcb = 10;
1797	TEST_RET_OLD(size_t, "%zx", pcb, ASMAtomicAddZ, 1, 11);
1798	TEST_RET_OLD(size_t, "%zx", pcb, ASMAtomicAddZ, ~(size_t)1, 9);
1799	TEST_RET_OLD(size_t, "%zx", pcb, ASMAtomicAddZ, ~(size_t)8, 0);
1800
1801	TEST_RET_OLD(size_t, "%zx", pcb, ASMAtomicSubZ, 0, 0);
1802	TEST_RET_OLD(size_t, "%zx", pcb, ASMAtomicSubZ, 10, ~(size_t)9);
1803	}
1804
1805	static void tstASMAtomicAdd(void)
1806	{
1807	DO_SIMPLE_TEST(ASMAtomicAddU32, uint32_t);
1808	DO_SIMPLE_TEST(ASMAtomicAddS32, int32_t);
1809	DO_SIMPLE_TEST(ASMAtomicAddU64, uint64_t);
1810	DO_SIMPLE_TEST(ASMAtomicAddS64, int64_t);
1811	DO_SIMPLE_TEST(ASMAtomicAddZ, size_t);
1812	}
1813
1814
1815	#define TEST_RET_NEW_NV(a_Type, a_Fmt, a_pVar, a_Function, a_VarExpect) do { \
1816	a_Type uNewRet = a_Function(a_pVar); \
1817	if (RT_LIKELY( uNewRet == (a_VarExpect) && *(a_pVar) == (a_VarExpect) )) { } \
1818	else RTTestFailed(g_hTest, "%s, %d: FAILURE: %s(%s) -> " a_Fmt " and %s=" a_Fmt ", expected both " a_Fmt "\n", \
1819	__FUNCTION__, __LINE__, #a_Function, #a_pVar, uNewRet, #a_pVar, *(a_pVar), (a_VarExpect)); \
1820	} while (0)
1821
1822
1823	DECLINLINE(void) tstASMAtomicDecIncU32Worker(uint32_t volatile *pu32)
1824	{
1825	*pu32 = 3;
1826	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicDecU32, 2);
1827	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicDecU32, 1);
1828	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicDecU32, 0);
1829	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicDecU32, UINT32_MAX);
1830	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicDecU32, UINT32_MAX - 1);
1831	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicDecU32, UINT32_MAX - 2);
1832	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicIncU32, UINT32_MAX - 1);
1833	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicIncU32, UINT32_MAX);
1834	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicIncU32, 0);
1835	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicIncU32, 1);
1836	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicIncU32, 2);
1837	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicDecU32, 1);
1838	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicIncU32, 2);
1839	*pu32 = _1M;
1840	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicDecU32, _1M - 1);
1841	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicIncU32, _1M);
1842	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicIncU32, _1M + 1);
1843	}
1844
1845	DECLINLINE(void) tstASMAtomicUoDecIncU32Worker(uint32_t volatile *pu32)
1846	{
1847	*pu32 = 3;
1848	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoDecU32, 2);
1849	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoDecU32, 1);
1850	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoDecU32, 0);
1851	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoDecU32, UINT32_MAX);
1852	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoDecU32, UINT32_MAX - 1);
1853	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoDecU32, UINT32_MAX - 2);
1854	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoIncU32, UINT32_MAX - 1);
1855	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoIncU32, UINT32_MAX);
1856	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoIncU32, 0);
1857	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoIncU32, 1);
1858	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoIncU32, 2);
1859	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoDecU32, 1);
1860	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoIncU32, 2);
1861	*pu32 = _1M;
1862	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoDecU32, _1M - 1);
1863	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoIncU32, _1M);
1864	TEST_RET_NEW_NV(uint32_t, "%#x", pu32, ASMAtomicUoIncU32, _1M + 1);
1865	}
1866
1867
1868	DECLINLINE(void) tstASMAtomicDecIncS32Worker(int32_t volatile *pi32)
1869	{
1870	*pi32 = 10;
1871	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 9);
1872	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 8);
1873	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 7);
1874	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 6);
1875	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 5);
1876	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 4);
1877	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 3);
1878	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 2);
1879	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 1);
1880	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 0);
1881	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, -1);
1882	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, -2);
1883	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicIncS32, -1);
1884	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicIncS32, 0);
1885	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicIncS32, 1);
1886	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicIncS32, 2);
1887	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicIncS32, 3);
1888	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 2);
1889	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicIncS32, 3);
1890	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, 2);
1891	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicIncS32, 3);
1892	*pi32 = INT32_MAX;
1893	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicDecS32, INT32_MAX - 1);
1894	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicIncS32, INT32_MAX);
1895	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicIncS32, INT32_MIN);
1896	}
1897
1898
1899	#if 0
1900	DECLINLINE(void) tstASMAtomicUoDecIncS32Worker(int32_t volatile *pi32)
1901	{
1902	*pi32 = 10;
1903	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 9);
1904	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 8);
1905	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 7);
1906	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 6);
1907	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 5);
1908	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 4);
1909	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 3);
1910	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 2);
1911	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 1);
1912	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 0);
1913	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, -1);
1914	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, -2);
1915	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoIncS32, -1);
1916	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoIncS32, 0);
1917	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoIncS32, 1);
1918	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoIncS32, 2);
1919	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoIncS32, 3);
1920	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 2);
1921	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoIncS32, 3);
1922	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, 2);
1923	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoIncS32, 3);
1924	*pi32 = INT32_MAX;
1925	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoDecS32, INT32_MAX - 1);
1926	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoIncS32, INT32_MAX);
1927	TEST_RET_NEW_NV(int32_t, "%d", pi32, ASMAtomicUoIncS32, INT32_MIN);
1928	}
1929	#endif
1930
1931
1932	DECLINLINE(void) tstASMAtomicDecIncU64Worker(uint64_t volatile *pu64)
1933	{
1934	*pu64 = 3;
1935	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicDecU64, 2);
1936	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicDecU64, 1);
1937	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicDecU64, 0);
1938	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicDecU64, UINT64_MAX);
1939	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicDecU64, UINT64_MAX - 1);
1940	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicDecU64, UINT64_MAX - 2);
1941	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicIncU64, UINT64_MAX - 1);
1942	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicIncU64, UINT64_MAX);
1943	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicIncU64, 0);
1944	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicIncU64, 1);
1945	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicIncU64, 2);
1946	*pu64 = _4G - 1;
1947	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicDecU64, _4G - 2);
1948	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicIncU64, _4G - 1);
1949	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicIncU64, _4G);
1950	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicIncU64, _4G + 1);
1951	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicDecU64, _4G);
1952	}
1953
1954
1955	#if 0
1956	DECLINLINE(void) tstASMAtomicUoDecIncU64Worker(uint64_t volatile *pu64)
1957	{
1958	*pu64 = 3;
1959	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoDecU64, 2);
1960	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoDecU64, 1);
1961	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoDecU64, 0);
1962	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoDecU64, UINT64_MAX);
1963	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoDecU64, UINT64_MAX - 1);
1964	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoDecU64, UINT64_MAX - 2);
1965	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoIncU64, UINT64_MAX - 1);
1966	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoIncU64, UINT64_MAX);
1967	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoIncU64, 0);
1968	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoIncU64, 1);
1969	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoIncU64, 2);
1970	*pu64 = _4G - 1;
1971	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoDecU64, _4G - 2);
1972	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoIncU64, _4G - 1);
1973	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoIncU64, _4G);
1974	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoIncU64, _4G + 1);
1975	TEST_RET_NEW_NV(uint64_t, "%lld", pu64, ASMAtomicUoDecU64, _4G);
1976	}
1977	#endif
1978
1979
1980	DECLINLINE(void) tstASMAtomicDecIncS64Worker(int64_t volatile *pi64)
1981	{
1982	*pi64 = 10;
1983	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 9);
1984	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 8);
1985	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 7);
1986	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 6);
1987	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 5);
1988	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 4);
1989	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 3);
1990	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 2);
1991	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 1);
1992	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 0);
1993	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, -1);
1994	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, -2);
1995	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicIncS64, -1);
1996	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicIncS64, 0);
1997	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicIncS64, 1);
1998	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicIncS64, 2);
1999	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicIncS64, 3);
2000	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 2);
2001	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicIncS64, 3);
2002	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, 2);
2003	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicIncS64, 3);
2004	*pi64 = INT64_MAX;
2005	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicDecS64, INT64_MAX - 1);
2006	}
2007
2008
2009	#if 0
2010	DECLINLINE(void) tstASMAtomicUoDecIncS64Worker(int64_t volatile *pi64)
2011	{
2012	*pi64 = 10;
2013	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 9);
2014	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 8);
2015	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 7);
2016	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 6);
2017	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 5);
2018	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 4);
2019	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 3);
2020	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 2);
2021	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 1);
2022	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 0);
2023	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, -1);
2024	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, -2);
2025	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoIncS64, -1);
2026	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoIncS64, 0);
2027	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoIncS64, 1);
2028	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoIncS64, 2);
2029	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoIncS64, 3);
2030	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 2);
2031	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoIncS64, 3);
2032	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, 2);
2033	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoIncS64, 3);
2034	*pi64 = INT64_MAX;
2035	TEST_RET_NEW_NV(int64_t, "%lld", pi64, ASMAtomicUoDecS64, INT64_MAX - 1);
2036	}
2037	#endif
2038
2039
2040	DECLINLINE(void) tstASMAtomicDecIncZWorker(size_t volatile *pcb)
2041	{
2042	size_t const uBaseVal = ~(size_t)0 >> 7;
2043	*pcb = uBaseVal;
2044	TEST_RET_NEW_NV(size_t, "%zx", pcb, ASMAtomicDecZ, uBaseVal - 1);
2045	TEST_RET_NEW_NV(size_t, "%zx", pcb, ASMAtomicDecZ, uBaseVal - 2);
2046	TEST_RET_NEW_NV(size_t, "%zx", pcb, ASMAtomicDecZ, uBaseVal - 3);
2047	TEST_RET_NEW_NV(size_t, "%zx", pcb, ASMAtomicIncZ, uBaseVal - 2);
2048	TEST_RET_NEW_NV(size_t, "%zx", pcb, ASMAtomicIncZ, uBaseVal - 1);
2049	TEST_RET_NEW_NV(size_t, "%zx", pcb, ASMAtomicIncZ, uBaseVal);
2050	TEST_RET_NEW_NV(size_t, "%zx", pcb, ASMAtomicIncZ, uBaseVal + 1);
2051	TEST_RET_NEW_NV(size_t, "%zx", pcb, ASMAtomicDecZ, uBaseVal);
2052	TEST_RET_NEW_NV(size_t, "%zx", pcb, ASMAtomicDecZ, uBaseVal - 1);
2053	TEST_RET_NEW_NV(size_t, "%zx", pcb, ASMAtomicIncZ, uBaseVal);
2054	}
2055
2056
2057	static void tstASMAtomicDecInc(void)
2058	{
2059	DO_SIMPLE_TEST(ASMAtomicDecIncU32, uint32_t);
2060	DO_SIMPLE_TEST(ASMAtomicUoDecIncU32, uint32_t);
2061	DO_SIMPLE_TEST(ASMAtomicDecIncS32, int32_t);
2062	//DO_SIMPLE_TEST(ASMAtomicUoDecIncS32, int32_t);
2063	DO_SIMPLE_TEST(ASMAtomicDecIncU64, uint64_t);
2064	//DO_SIMPLE_TEST(ASMAtomicUoDecIncU64, uint64_t);
2065	DO_SIMPLE_TEST(ASMAtomicDecIncS64, int64_t);
2066	//DO_SIMPLE_TEST(ASMAtomicUoDecIncS64, int64_t);
2067	DO_SIMPLE_TEST(ASMAtomicDecIncZ, size_t);
2068	}
2069
2070
2071	#define TEST_RET_VOID(a_Type, a_Fmt, a_pVar, a_Function, a_uVal, a_VarExpect) do { \
2072	a_Function(a_pVar, a_uVal); \
2073	if (RT_LIKELY( *(a_pVar) == (a_VarExpect) )) { } \
2074	else RTTestFailed(g_hTest, "%s, %d: FAILURE: %s(%s, " a_Fmt ") -> %s=" a_Fmt ", expected " a_Fmt "\n", \
2075	__FUNCTION__, __LINE__, #a_Function, #a_pVar, a_uVal, #a_pVar, *(a_pVar), (a_VarExpect)); \
2076	} while (0)
2077
2078	#define TEST_RET_NEW(a_Type, a_Fmt, a_pVar, a_Function, a_uVal, a_VarExpect) do { \
2079	a_Type uNewRet = a_Function(a_pVar, a_uVal); \
2080	if (RT_LIKELY( uNewRet == (a_VarExpect) && *(a_pVar) == (a_VarExpect) )) { } \
2081	else RTTestFailed(g_hTest, "%s, %d: FAILURE: %s(%s, " a_Fmt ") -> " a_Fmt " and %s=" a_Fmt ", expected both " a_Fmt "\n", \
2082	__FUNCTION__, __LINE__, #a_Function, #a_pVar, a_uVal, uNewRet, #a_pVar, *(a_pVar), (a_VarExpect)); \
2083	} while (0)
2084
2085
2086	DECLINLINE(void) tstASMAtomicAndOrXorU32Worker(uint32_t volatile *pu32)
2087	{
2088	*pu32 = UINT32_C(0xffffffff);
2089	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicOrU32, UINT32_C(0xffffffff), UINT32_C(0xffffffff));
2090	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicAndU32, UINT32_C(0xffffffff), UINT32_C(0xffffffff));
2091	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicAndU32, UINT32_C(0x8f8f8f8f), UINT32_C(0x8f8f8f8f));
2092	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicOrU32, UINT32_C(0x70707070), UINT32_C(0xffffffff));
2093	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicAndU32, UINT32_C(1), UINT32_C(1));
2094	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicOrU32, UINT32_C(0x80000000), UINT32_C(0x80000001));
2095	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicAndU32, UINT32_C(0x80000000), UINT32_C(0x80000000));
2096	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicAndU32, UINT32_C(0), UINT32_C(0));
2097	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicOrU32, UINT32_C(0x42424242), UINT32_C(0x42424242));
2098	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicAndU32, UINT32_C(0x00ff0f00), UINT32_C(0x00420200));
2099	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicXorU32, UINT32_C(0x42004042), UINT32_C(0x42424242));
2100	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicXorU32, UINT32_C(0xff024200), UINT32_C(0xbd400042));
2101	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicXorU32, UINT32_C(0x00000000), UINT32_C(0xbd400042));
2102	}
2103
2104
2105	DECLINLINE(void) tstASMAtomicUoAndOrXorU32Worker(uint32_t volatile *pu32)
2106	{
2107	*pu32 = UINT32_C(0xffffffff);
2108	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoOrU32, UINT32_C(0xffffffff), UINT32_C(0xffffffff));
2109	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoAndU32, UINT32_C(0xffffffff), UINT32_C(0xffffffff));
2110	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoAndU32, UINT32_C(0x8f8f8f8f), UINT32_C(0x8f8f8f8f));
2111	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoOrU32, UINT32_C(0x70707070), UINT32_C(0xffffffff));
2112	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoAndU32, UINT32_C(1), UINT32_C(1));
2113	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoOrU32, UINT32_C(0x80000000), UINT32_C(0x80000001));
2114	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoAndU32, UINT32_C(0x80000000), UINT32_C(0x80000000));
2115	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoAndU32, UINT32_C(0), UINT32_C(0));
2116	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoOrU32, UINT32_C(0x42424242), UINT32_C(0x42424242));
2117	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoAndU32, UINT32_C(0x00ff0f00), UINT32_C(0x00420200));
2118	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoXorU32, UINT32_C(0x42004042), UINT32_C(0x42424242));
2119	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoXorU32, UINT32_C(0xff024200), UINT32_C(0xbd400042));
2120	TEST_RET_VOID(uint32_t, "%#x", pu32, ASMAtomicUoXorU32, UINT32_C(0x00000000), UINT32_C(0xbd400042));
2121	}
2122
2123
2124	DECLINLINE(void) tstASMAtomicAndOrXorExU32Worker(uint32_t volatile *pu32)
2125	{
2126	*pu32 = UINT32_C(0xffffffff);
2127	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicOrExU32, UINT32_C(0xffffffff), UINT32_C(0xffffffff));
2128	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAndExU32, UINT32_C(0xffffffff), UINT32_C(0xffffffff));
2129	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAndExU32, UINT32_C(0x8f8f8f8f), UINT32_C(0x8f8f8f8f));
2130	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicOrExU32, UINT32_C(0x70707070), UINT32_C(0xffffffff));
2131	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAndExU32, UINT32_C(1), UINT32_C(1));
2132	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicOrExU32, UINT32_C(0x80000000), UINT32_C(0x80000001));
2133	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAndExU32, UINT32_C(0x80000000), UINT32_C(0x80000000));
2134	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAndExU32, UINT32_C(0), UINT32_C(0));
2135	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicOrExU32, UINT32_C(0x42424242), UINT32_C(0x42424242));
2136	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicAndExU32, UINT32_C(0x00ff0f00), UINT32_C(0x00420200));
2137	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicXorExU32, UINT32_C(0x42004042), UINT32_C(0x42424242));
2138	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicXorExU32, UINT32_C(0xff024200), UINT32_C(0xbd400042));
2139	TEST_RET_OLD(uint32_t, "%#x", pu32, ASMAtomicXorExU32, UINT32_C(0x00000000), UINT32_C(0xbd400042));
2140	}
2141
2142
2143	DECLINLINE(void) tstASMAtomicAndOrXorU64Worker(uint64_t volatile *pu64)
2144	{
2145	*pu64 = UINT64_C(0xffffffff);
2146	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicOrU64, UINT64_C(0xffffffff), UINT64_C(0xffffffff));
2147	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0xffffffff), UINT64_C(0xffffffff));
2148	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0x8f8f8f8f), UINT64_C(0x8f8f8f8f));
2149	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicOrU64, UINT64_C(0x70707070), UINT64_C(0xffffffff));
2150	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(1), UINT64_C(1));
2151	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicOrU64, UINT64_C(0x80000000), UINT64_C(0x80000001));
2152	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0x80000000), UINT64_C(0x80000000));
2153	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0), UINT64_C(0));
2154	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicOrU64, UINT64_C(0x42424242), UINT64_C(0x42424242));
2155	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0x00ff0f00), UINT64_C(0x00420200));
2156	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicXorU64, UINT64_C(0x42004042), UINT64_C(0x42424242));
2157	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicXorU64, UINT64_C(0xff024200), UINT64_C(0xbd400042));
2158	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicXorU64, UINT64_C(0x00000000), UINT64_C(0xbd400042));
2159
2160	/* full 64-bit */
2161	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0x0000000000000000), UINT64_C(0x0000000000000000));
2162	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicOrU64, UINT64_C(0xffffffffffffffff), UINT64_C(0xffffffffffffffff));
2163	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0xffffffffffffffff), UINT64_C(0xffffffffffffffff));
2164	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0x8f8f8f8f8f8f8f8f), UINT64_C(0x8f8f8f8f8f8f8f8f));
2165	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicOrU64, UINT64_C(0x7070707070707070), UINT64_C(0xffffffffffffffff));
2166	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0x0000000000000001), UINT64_C(0x0000000000000001));
2167	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicOrU64, UINT64_C(0x8000000000000000), UINT64_C(0x8000000000000001));
2168	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0x8000000000000000), UINT64_C(0x8000000000000000));
2169	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0), UINT64_C(0));
2170	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicOrU64, UINT64_C(0x4242424242424242), UINT64_C(0x4242424242424242));
2171	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicAndU64, UINT64_C(0x00ff0f00ff0f0000), UINT64_C(0x0042020042020000));
2172	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicXorU64, UINT64_C(0x4200404242040000), UINT64_C(0x4242424242420000));
2173	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicXorU64, UINT64_C(0xff02420000ff2127), UINT64_C(0xbd40004242bd2127));
2174	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicXorU64, UINT64_C(0x0000000000000000), UINT64_C(0xbd40004242bd2127));
2175	}
2176
2177
2178	DECLINLINE(void) tstASMAtomicUoAndOrXorU64Worker(uint64_t volatile *pu64)
2179	{
2180	*pu64 = UINT64_C(0xffffffff);
2181	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoOrU64, UINT64_C(0xffffffff), UINT64_C(0xffffffff));
2182	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0xffffffff), UINT64_C(0xffffffff));
2183	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0x8f8f8f8f), UINT64_C(0x8f8f8f8f));
2184	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoOrU64, UINT64_C(0x70707070), UINT64_C(0xffffffff));
2185	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(1), UINT64_C(1));
2186	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoOrU64, UINT64_C(0x80000000), UINT64_C(0x80000001));
2187	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0x80000000), UINT64_C(0x80000000));
2188	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0), UINT64_C(0));
2189	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoOrU64, UINT64_C(0x42424242), UINT64_C(0x42424242));
2190	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0x00ff0f00), UINT64_C(0x00420200));
2191	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoXorU64, UINT64_C(0x42004042), UINT64_C(0x42424242));
2192	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoXorU64, UINT64_C(0xff024200), UINT64_C(0xbd400042));
2193	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoXorU64, UINT64_C(0x00000000), UINT64_C(0xbd400042));
2194
2195	/* full 64-bit */
2196	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0x0000000000000000), UINT64_C(0x0000000000000000));
2197	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoOrU64, UINT64_C(0xffffffffffffffff), UINT64_C(0xffffffffffffffff));
2198	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0xffffffffffffffff), UINT64_C(0xffffffffffffffff));
2199	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0x8f8f8f8f8f8f8f8f), UINT64_C(0x8f8f8f8f8f8f8f8f));
2200	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoOrU64, UINT64_C(0x7070707070707070), UINT64_C(0xffffffffffffffff));
2201	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0x0000000000000001), UINT64_C(0x0000000000000001));
2202	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoOrU64, UINT64_C(0x8000000000000000), UINT64_C(0x8000000000000001));
2203	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0x8000000000000000), UINT64_C(0x8000000000000000));
2204	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0), UINT64_C(0));
2205	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoOrU64, UINT64_C(0x4242424242424242), UINT64_C(0x4242424242424242));
2206	TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoAndU64, UINT64_C(0x00ff0f00ff0f0000), UINT64_C(0x0042020042020000));
2207	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoXorU64, UINT64_C(0x4200404242040000), UINT64_C(0x4242424242420000));
2208	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoXorU64, UINT64_C(0xff02420000ff2127), UINT64_C(0xbd40004242bd2127));
2209	//TEST_RET_VOID(uint64_t, "%#llx", pu64, ASMAtomicUoXorU64, UINT64_C(0x0000000000000000), UINT64_C(0xbd40004242bd2127));
2210	}
2211
2212
2213	#if 0
2214	DECLINLINE(void) tstASMAtomicAndOrXorExU64Worker(uint64_t volatile *pu64)
2215	{
2216	*pu64 = UINT64_C(0xffffffff);
2217	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicOrExU64, UINT64_C(0xffffffff), UINT64_C(0xffffffff));
2218	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0xffffffff), UINT64_C(0xffffffff));
2219	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0x8f8f8f8f), UINT64_C(0x8f8f8f8f));
2220	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicOrExU64, UINT64_C(0x70707070), UINT64_C(0xffffffff));
2221	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(1), UINT64_C(1));
2222	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicOrExU64, UINT64_C(0x80000000), UINT64_C(0x80000001));
2223	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0x80000000), UINT64_C(0x80000000));
2224	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0), UINT64_C(0));
2225	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicOrExU64, UINT64_C(0x42424242), UINT64_C(0x42424242));
2226	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0x00ff0f00), UINT64_C(0x00420200));
2227	//TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicXorExU64, UINT64_C(0x42004042), UINT64_C(0x42424242));
2228	//TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicXorExU64, UINT64_C(0xff024200), UINT64_C(0xbd400042));
2229	//TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicXorExU64, UINT64_C(0x00000000), UINT64_C(0xbd400042));
2230
2231	/* full 64-bit */
2232	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0x0000000000000000), UINT64_C(0x0000000000000000));
2233	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicOrExU64, UINT64_C(0xffffffffffffffff), UINT64_C(0xffffffffffffffff));
2234	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0xffffffffffffffff), UINT64_C(0xffffffffffffffff));
2235	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0x8f8f8f8f8f8f8f8f), UINT64_C(0x8f8f8f8f8f8f8f8f));
2236	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicOrExU64, UINT64_C(0x7070707070707070), UINT64_C(0xffffffffffffffff));
2237	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0x0000000000000001), UINT64_C(0x0000000000000001));
2238	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicOrExU64, UINT64_C(0x8000000000000000), UINT64_C(0x8000000000000001));
2239	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0x8000000000000000), UINT64_C(0x8000000000000000));
2240	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0), UINT64_C(0));
2241	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicOrExU64, UINT64_C(0x4242424242424242), UINT64_C(0x4242424242424242));
2242	TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicAndExU64, UINT64_C(0x00ff0f00ff0f0000), UINT64_C(0x0042020042020000));
2243	//TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicXorExU64, UINT64_C(0x4200404242040000), UINT64_C(0x4242424242420000));
2244	//TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicXorExU64, UINT64_C(0xff02420000ff2127), UINT64_C(0xbd40004242bd2127));
2245	//TEST_RET_OLD(uint64_t, "%#llx", pu64, ASMAtomicXorExU64, UINT64_C(0x0000000000000000), UINT64_C(0xbd40004242bd2127));
2246	}
2247	#endif
2248
2249
2250	static void tstASMAtomicAndOrXor(void)
2251	{
2252	DO_SIMPLE_TEST(ASMAtomicAndOrXorU32, uint32_t);
2253	DO_SIMPLE_TEST(ASMAtomicUoAndOrXorU32, uint32_t);
2254	DO_SIMPLE_TEST(ASMAtomicAndOrXorExU32, uint32_t);
2255	DO_SIMPLE_TEST(ASMAtomicAndOrXorU64, uint64_t);
2256	DO_SIMPLE_TEST(ASMAtomicUoAndOrXorU64, uint64_t);
2257	//DO_SIMPLE_TEST(ASMAtomicAndOrXorExU64, uint64_t);
2258	}
2259
2260
2261	typedef struct
2262	{
2263	uint8_t ab[PAGE_SIZE];
2264	} TSTPAGE;
2265
2266
2267	DECLINLINE(void) tstASMMemZeroPageWorker(TSTPAGE *pPage)
2268	{
2269	for (unsigned j = 0; j < 16; j++)
2270	{
2271	memset(pPage, 0x11 * j, sizeof(*pPage));
2272	ASMMemZeroPage(pPage);
2273	for (unsigned i = 0; i < sizeof(pPage->ab); i++)
2274	if (pPage->ab[i])
2275	RTTestFailed(g_hTest, "ASMMemZeroPage didn't clear byte at offset %#x!\n", i);
2276	if (ASMMemIsZeroPage(pPage) != true)
2277	RTTestFailed(g_hTest, "ASMMemIsZeroPage returns false after ASMMemZeroPage!\n");
2278	if (ASMMemFirstMismatchingU32(pPage, sizeof(pPage), 0) != NULL)
2279	RTTestFailed(g_hTest, "ASMMemFirstMismatchingU32(,,0) returns non-NULL after ASMMemZeroPage!\n");
2280	}
2281	}
2282
2283
2284	static void tstASMMemZeroPage(void)
2285	{
2286	RTTestISub("ASMMemZeroPage");
2287	DO_SIMPLE_TEST_NO_SUB_NO_STACK(tstASMMemZeroPageWorker, TSTPAGE);
2288	}
2289
2290
2291	void tstASMMemIsZeroPage(RTTEST hTest)
2292	{
2293	RTTestSub(hTest, "ASMMemIsZeroPage");
2294
2295	void *pvPage1 = RTTestGuardedAllocHead(hTest, PAGE_SIZE);
2296	void *pvPage2 = RTTestGuardedAllocTail(hTest, PAGE_SIZE);
2297	RTTESTI_CHECK_RETV(pvPage1 && pvPage2);
2298
2299	memset(pvPage1, 0, PAGE_SIZE);
2300	memset(pvPage2, 0, PAGE_SIZE);
2301	RTTESTI_CHECK(ASMMemIsZeroPage(pvPage1));
2302	RTTESTI_CHECK(ASMMemIsZeroPage(pvPage2));
2303
2304	memset(pvPage1, 0xff, PAGE_SIZE);
2305	memset(pvPage2, 0xff, PAGE_SIZE);
2306	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage1));
2307	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage2));
2308
2309	memset(pvPage1, 0, PAGE_SIZE);
2310	memset(pvPage2, 0, PAGE_SIZE);
2311	for (unsigned off = 0; off < PAGE_SIZE; off++)
2312	{
2313	((uint8_t *)pvPage1)[off] = 1;
2314	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage1));
2315	((uint8_t *)pvPage1)[off] = 0;
2316
2317	((uint8_t *)pvPage2)[off] = 0x80;
2318	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage2));
2319	((uint8_t *)pvPage2)[off] = 0;
2320	}
2321
2322	RTTestSubDone(hTest);
2323	}
2324
2325
2326	void tstASMMemFirstMismatchingU8(RTTEST hTest)
2327	{
2328	RTTestSub(hTest, "ASMMemFirstMismatchingU8");
2329
2330	uint8_t pbPage1 = (uint8_t )RTTestGuardedAllocHead(hTest, PAGE_SIZE);
2331	uint8_t pbPage2 = (uint8_t )RTTestGuardedAllocTail(hTest, PAGE_SIZE);
2332	RTTESTI_CHECK_RETV(pbPage1 && pbPage2);
2333
2334	memset(pbPage1, 0, PAGE_SIZE);
2335	memset(pbPage2, 0, PAGE_SIZE);
2336	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0) == NULL);
2337	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0) == NULL);
2338	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 1) == pbPage1);
2339	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 1) == pbPage2);
2340	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0x87) == pbPage1);
2341	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0x87) == pbPage2);
2342	RTTESTI_CHECK(ASMMemIsZero(pbPage1, PAGE_SIZE));
2343	RTTESTI_CHECK(ASMMemIsZero(pbPage2, PAGE_SIZE));
2344	RTTESTI_CHECK(ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0));
2345	RTTESTI_CHECK(ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0));
2346	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0x34));
2347	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0x88));
2348	unsigned cbSub = 32;
2349	while (cbSub-- > 0)
2350	{
2351	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0) == NULL);
2352	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0) == NULL);
2353	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0) == NULL);
2354	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0) == NULL);
2355
2356	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0x34) == &pbPage1[PAGE_SIZE - cbSub] \|\| !cbSub);
2357	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0x99) == &pbPage2[PAGE_SIZE - cbSub] \|\| !cbSub);
2358	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0x42) == pbPage1 \|\| !cbSub);
2359	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0x88) == pbPage2 \|\| !cbSub);
2360	}
2361
2362	memset(pbPage1, 0xff, PAGE_SIZE);
2363	memset(pbPage2, 0xff, PAGE_SIZE);
2364	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0xff) == NULL);
2365	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0xff) == NULL);
2366	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0xfe) == pbPage1);
2367	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0xfe) == pbPage2);
2368	RTTESTI_CHECK(!ASMMemIsZero(pbPage1, PAGE_SIZE));
2369	RTTESTI_CHECK(!ASMMemIsZero(pbPage2, PAGE_SIZE));
2370	RTTESTI_CHECK(ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0xff));
2371	RTTESTI_CHECK(ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0xff));
2372	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0));
2373	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0));
2374	cbSub = 32;
2375	while (cbSub-- > 0)
2376	{
2377	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0xff) == NULL);
2378	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0xff) == NULL);
2379	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0xff) == NULL);
2380	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0xff) == NULL);
2381
2382	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0xfe) == &pbPage1[PAGE_SIZE - cbSub] \|\| !cbSub);
2383	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0xfe) == &pbPage2[PAGE_SIZE - cbSub] \|\| !cbSub);
2384	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0xfe) == pbPage1 \|\| !cbSub);
2385	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0xfe) == pbPage2 \|\| !cbSub);
2386	}
2387
2388
2389	/*
2390	* Various alignments and sizes.
2391	*/
2392	uint8_t const bFiller1 = 0x00;
2393	uint8_t const bFiller2 = 0xf6;
2394	size_t const cbBuf = 128;
2395	uint8_t *pbBuf1 = pbPage1;
2396	uint8_t pbBuf2 = &pbPage2[PAGE_SIZE - cbBuf]; / Put it up against the tail guard */
2397	memset(pbPage1, ~bFiller1, PAGE_SIZE);
2398	memset(pbPage2, ~bFiller2, PAGE_SIZE);
2399	memset(pbBuf1, bFiller1, cbBuf);
2400	memset(pbBuf2, bFiller2, cbBuf);
2401	for (size_t offNonZero = 0; offNonZero < cbBuf; offNonZero++)
2402	{
2403	uint8_t bRand = (uint8_t)RTRandU32();
2404	pbBuf1[offNonZero] = bRand \| 1;
2405	pbBuf2[offNonZero] = (0x80 \| bRand) ^ 0xf6;
2406
2407	for (size_t offStart = 0; offStart < 32; offStart++)
2408	{
2409	size_t const cbMax = cbBuf - offStart;
2410	for (size_t cb = 0; cb < cbMax; cb++)
2411	{
2412	size_t const offEnd = offStart + cb;
2413	uint8_t bSaved1, bSaved2;
2414	if (offEnd < PAGE_SIZE)
2415	{
2416	bSaved1 = pbBuf1[offEnd];
2417	bSaved2 = pbBuf2[offEnd];
2418	pbBuf1[offEnd] = 0xff;
2419	pbBuf2[offEnd] = 0xff;
2420	}
2421	#ifdef _MSC_VER /* simple stupid compiler warnings */
2422	else
2423	bSaved1 = bSaved2 = 0;
2424	#endif
2425
2426	uint8_t pbRet = (uint8_t )ASMMemFirstMismatchingU8(pbBuf1 + offStart, cb, bFiller1);
2427	RTTESTI_CHECK(offNonZero - offStart < cb ? pbRet == &pbBuf1[offNonZero] : pbRet == NULL);
2428
2429	pbRet = (uint8_t *)ASMMemFirstMismatchingU8(pbBuf2 + offStart, cb, bFiller2);
2430	RTTESTI_CHECK(offNonZero - offStart < cb ? pbRet == &pbBuf2[offNonZero] : pbRet == NULL);
2431
2432	if (offEnd < PAGE_SIZE)
2433	{
2434	pbBuf1[offEnd] = bSaved1;
2435	pbBuf2[offEnd] = bSaved2;
2436	}
2437	}
2438	}
2439
2440	pbBuf1[offNonZero] = 0;
2441	pbBuf2[offNonZero] = 0xf6;
2442	}
2443
2444	RTTestSubDone(hTest);
2445	}
2446
2447
2448	typedef struct TSTBUF32 { uint32_t au32[384]; } TSTBUF32;
2449
2450	DECLINLINE(void) tstASMMemZero32Worker(TSTBUF32 *pBuf)
2451	{
2452	ASMMemZero32(pBuf, sizeof(*pBuf));
2453	for (unsigned i = 0; i < RT_ELEMENTS(pBuf->au32); i++)
2454	if (pBuf->au32[i])
2455	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear dword at index %#x!\n", i);
2456	if (ASMMemFirstNonZero(pBuf, sizeof(*pBuf)) != NULL)
2457	RTTestFailed(g_hTest, "ASMMemFirstNonZero return non-NULL after ASMMemZero32\n");
2458	if (!ASMMemIsZero(pBuf, sizeof(*pBuf)))
2459	RTTestFailed(g_hTest, "ASMMemIsZero return false after ASMMemZero32\n");
2460
2461	memset(pBuf, 0xfe, sizeof(*pBuf));
2462	ASMMemZero32(pBuf, sizeof(*pBuf));
2463	for (unsigned i = 0; i < RT_ELEMENTS(pBuf->au32); i++)
2464	if (pBuf->au32[i])
2465	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear dword at index %#x!\n", i);
2466	if (ASMMemFirstNonZero(pBuf, sizeof(*pBuf)) != NULL)
2467	RTTestFailed(g_hTest, "ASMMemFirstNonZero return non-NULL after ASMMemZero32\n");
2468	if (!ASMMemIsZero(pBuf, sizeof(*pBuf)))
2469	RTTestFailed(g_hTest, "ASMMemIsZero return false after ASMMemZero32\n");
2470	}
2471
2472
2473	void tstASMMemZero32(void)
2474	{
2475	RTTestSub(g_hTest, "ASMMemZero32");
2476
2477	struct
2478	{
2479	uint64_t u64Magic1;
2480	uint8_t abPage[PAGE_SIZE - 32];
2481	uint64_t u64Magic2;
2482	} Buf1, Buf2, Buf3;
2483
2484	Buf1.u64Magic1 = UINT64_C(0xffffffffffffffff);
2485	memset(Buf1.abPage, 0x55, sizeof(Buf1.abPage));
2486	Buf1.u64Magic2 = UINT64_C(0xffffffffffffffff);
2487	Buf2.u64Magic1 = UINT64_C(0xffffffffffffffff);
2488	memset(Buf2.abPage, 0x77, sizeof(Buf2.abPage));
2489	Buf2.u64Magic2 = UINT64_C(0xffffffffffffffff);
2490	Buf3.u64Magic1 = UINT64_C(0xffffffffffffffff);
2491	memset(Buf3.abPage, 0x99, sizeof(Buf3.abPage));
2492	Buf3.u64Magic2 = UINT64_C(0xffffffffffffffff);
2493	ASMMemZero32(Buf1.abPage, sizeof(Buf1.abPage));
2494	ASMMemZero32(Buf2.abPage, sizeof(Buf2.abPage));
2495	ASMMemZero32(Buf3.abPage, sizeof(Buf3.abPage));
2496	if ( Buf1.u64Magic1 != UINT64_C(0xffffffffffffffff)
2497	\|\| Buf1.u64Magic2 != UINT64_C(0xffffffffffffffff)
2498	\|\| Buf2.u64Magic1 != UINT64_C(0xffffffffffffffff)
2499	\|\| Buf2.u64Magic2 != UINT64_C(0xffffffffffffffff)
2500	\|\| Buf3.u64Magic1 != UINT64_C(0xffffffffffffffff)
2501	\|\| Buf3.u64Magic2 != UINT64_C(0xffffffffffffffff))
2502	{
2503	RTTestFailed(g_hTest, "ASMMemZero32 violated one/both magic(s)!\n");
2504	}
2505	for (unsigned i = 0; i < RT_ELEMENTS(Buf1.abPage); i++)
2506	if (Buf1.abPage[i])
2507	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
2508	for (unsigned i = 0; i < RT_ELEMENTS(Buf2.abPage); i++)
2509	if (Buf2.abPage[i])
2510	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
2511	for (unsigned i = 0; i < RT_ELEMENTS(Buf3.abPage); i++)
2512	if (Buf3.abPage[i])
2513	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
2514
2515	DO_SIMPLE_TEST_NO_SUB(tstASMMemZero32Worker, TSTBUF32);
2516	}
2517
2518
2519	DECLINLINE(void) tstASMMemFill32Worker(TSTBUF32 *pBuf)
2520	{
2521	ASMMemFill32(pBuf, sizeof(*pBuf), UINT32_C(0xf629bce1));
2522	for (unsigned i = 0; i < RT_ELEMENTS(pBuf->au32); i++)
2523	if (pBuf->au32[i] != UINT32_C(0xf629bce1))
2524	RTTestFailed(g_hTest, "ASMMemFill32 didn't set dword at index %#x correctly!\n", i);
2525	if (ASMMemFirstMismatchingU32(pBuf, sizeof(*pBuf), UINT32_C(0xf629bce1)) != NULL)
2526	RTTestFailed(g_hTest, "ASMMemFirstMismatchingU32(,,UINT32_C(0xf629bce1)) returns non-NULL after ASMMemFill32!\n");
2527
2528	memset(pBuf, 0xfe, sizeof(*pBuf));
2529	ASMMemFill32(pBuf, sizeof(*pBuf), UINT32_C(0x12345678));
2530	for (unsigned i = 0; i < RT_ELEMENTS(pBuf->au32); i++)
2531	if (pBuf->au32[i] != UINT32_C(0x12345678))
2532	RTTestFailed(g_hTest, "ASMMemFill32 didn't set dword at index %#x correctly!\n", i);
2533	if (ASMMemFirstMismatchingU32(pBuf, sizeof(*pBuf), UINT32_C(0x12345678)) != NULL)
2534	RTTestFailed(g_hTest, "ASMMemFirstMismatchingU32(,,UINT32_C(0x12345678)) returns non-NULL after ASMMemFill32!\n");
2535	}
2536
2537	void tstASMMemFill32(void)
2538	{
2539	RTTestSub(g_hTest, "ASMMemFill32");
2540
2541	struct
2542	{
2543	uint64_t u64Magic1;
2544	uint32_t au32Page[PAGE_SIZE / 4];
2545	uint64_t u64Magic2;
2546	} Buf1;
2547	struct
2548	{
2549	uint64_t u64Magic1;
2550	uint32_t au32Page[(PAGE_SIZE / 4) - 3];
2551	uint64_t u64Magic2;
2552	} Buf2;
2553	struct
2554	{
2555	uint64_t u64Magic1;
2556	uint32_t au32Page[(PAGE_SIZE / 4) - 1];
2557	uint64_t u64Magic2;
2558	} Buf3;
2559
2560	Buf1.u64Magic1 = UINT64_C(0xffffffffffffffff);
2561	memset(Buf1.au32Page, 0x55, sizeof(Buf1.au32Page));
2562	Buf1.u64Magic2 = UINT64_C(0xffffffffffffffff);
2563	Buf2.u64Magic1 = UINT64_C(0xffffffffffffffff);
2564	memset(Buf2.au32Page, 0x77, sizeof(Buf2.au32Page));
2565	Buf2.u64Magic2 = UINT64_C(0xffffffffffffffff);
2566	Buf3.u64Magic1 = UINT64_C(0xffffffffffffffff);
2567	memset(Buf3.au32Page, 0x99, sizeof(Buf3.au32Page));
2568	Buf3.u64Magic2 = UINT64_C(0xffffffffffffffff);
2569	ASMMemFill32(Buf1.au32Page, sizeof(Buf1.au32Page), 0xdeadbeef);
2570	ASMMemFill32(Buf2.au32Page, sizeof(Buf2.au32Page), 0xcafeff01);
2571	ASMMemFill32(Buf3.au32Page, sizeof(Buf3.au32Page), 0xf00dd00f);
2572	if ( Buf1.u64Magic1 != UINT64_C(0xffffffffffffffff)
2573	\|\| Buf1.u64Magic2 != UINT64_C(0xffffffffffffffff)
2574	\|\| Buf2.u64Magic1 != UINT64_C(0xffffffffffffffff)
2575	\|\| Buf2.u64Magic2 != UINT64_C(0xffffffffffffffff)
2576	\|\| Buf3.u64Magic1 != UINT64_C(0xffffffffffffffff)
2577	\|\| Buf3.u64Magic2 != UINT64_C(0xffffffffffffffff))
2578	RTTestFailed(g_hTest, "ASMMemFill32 violated one/both magic(s)!\n");
2579	for (unsigned i = 0; i < RT_ELEMENTS(Buf1.au32Page); i++)
2580	if (Buf1.au32Page[i] != 0xdeadbeef)
2581	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf1.au32Page[i], 0xdeadbeef);
2582	for (unsigned i = 0; i < RT_ELEMENTS(Buf2.au32Page); i++)
2583	if (Buf2.au32Page[i] != 0xcafeff01)
2584	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf2.au32Page[i], 0xcafeff01);
2585	for (unsigned i = 0; i < RT_ELEMENTS(Buf3.au32Page); i++)
2586	if (Buf3.au32Page[i] != 0xf00dd00f)
2587	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf3.au32Page[i], 0xf00dd00f);
2588
2589	DO_SIMPLE_TEST_NO_SUB(tstASMMemFill32Worker, TSTBUF32);
2590	}
2591
2592
2593	void tstASMProbe(RTTEST hTest)
2594	{
2595	RTTestSub(hTest, "ASMProbeReadByte/Buffer");
2596
2597	uint8_t b = 42;
2598	RTTESTI_CHECK(ASMProbeReadByte(&b) == 42);
2599	ASMProbeReadBuffer(&b, sizeof(b));
2600
2601	for (uint32_t cPages = 1; cPages < 16; cPages++)
2602	{
2603	uint8_t pbBuf1 = (uint8_t )RTTestGuardedAllocHead(hTest, cPages * PAGE_SIZE);
2604	uint8_t pbBuf2 = (uint8_t )RTTestGuardedAllocTail(hTest, cPages * PAGE_SIZE);
2605	RTTESTI_CHECK_RETV(pbBuf1 && pbBuf2);
2606
2607	memset(pbBuf1, 0xf6, cPages * PAGE_SIZE);
2608	memset(pbBuf2, 0x42, cPages * PAGE_SIZE);
2609
2610	RTTESTI_CHECK(ASMProbeReadByte(&pbBuf1[cPages * PAGE_SIZE - 1]) == 0xf6);
2611	RTTESTI_CHECK(ASMProbeReadByte(&pbBuf2[cPages * PAGE_SIZE - 1]) == 0x42);
2612	RTTESTI_CHECK(ASMProbeReadByte(&pbBuf1[0]) == 0xf6);
2613	RTTESTI_CHECK(ASMProbeReadByte(&pbBuf2[0]) == 0x42);
2614
2615	ASMProbeReadBuffer(pbBuf1, cPages * PAGE_SIZE);
2616	ASMProbeReadBuffer(pbBuf2, cPages * PAGE_SIZE);
2617	}
2618	}
2619
2620
2621	void tstASMMisc(void)
2622	{
2623	RTTestSub(g_hTest, "Misc");
2624	for (uint32_t i = 0; i < 20; i++)
2625	{
2626	ASMWriteFence();
2627	ASMCompilerBarrier();
2628	ASMReadFence();
2629	ASMNopPause();
2630	ASMSerializeInstruction();
2631	ASMMemoryFence();
2632	}
2633	}
2634
2635	void tstASMMath(void)
2636	{
2637	RTTestSub(g_hTest, "Math");
2638
2639	uint64_t u64 = ASMMult2xU32RetU64(UINT32_C(0x80000000), UINT32_C(0x10000000));
2640	CHECKVAL(u64, UINT64_C(0x0800000000000000), "%#018RX64");
2641
2642	uint32_t u32 = ASMDivU64ByU32RetU32(UINT64_C(0x0800000000000000), UINT32_C(0x10000000));
2643	CHECKVAL(u32, UINT32_C(0x80000000), "%#010RX32");
2644
2645	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x00000001), UINT32_C(0x00000001), UINT32_C(0x00000001));
2646	CHECKVAL(u32, UINT32_C(0x00000001), "%#018RX32");
2647	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x10000000), UINT32_C(0x80000000), UINT32_C(0x20000000));
2648	CHECKVAL(u32, UINT32_C(0x40000000), "%#018RX32");
2649	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x76543210), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
2650	CHECKVAL(u32, UINT32_C(0x76543210), "%#018RX32");
2651	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0xffffffff), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
2652	CHECKVAL(u32, UINT32_C(0xffffffff), "%#018RX32");
2653	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0xffffffff), UINT32_C(0xfffffff0), UINT32_C(0xffffffff));
2654	CHECKVAL(u32, UINT32_C(0xfffffff0), "%#018RX32");
2655	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x10359583), UINT32_C(0x58734981), UINT32_C(0xf8694045));
2656	CHECKVAL(u32, UINT32_C(0x05c584ce), "%#018RX32");
2657	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x10359583), UINT32_C(0xf8694045), UINT32_C(0x58734981));
2658	CHECKVAL(u32, UINT32_C(0x2d860795), "%#018RX32");
2659
2660	#if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
2661	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x0000000000000001), UINT32_C(0x00000001), UINT32_C(0x00000001));
2662	CHECKVAL(u64, UINT64_C(0x0000000000000001), "%#018RX64");
2663	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x0000000100000000), UINT32_C(0x80000000), UINT32_C(0x00000002));
2664	CHECKVAL(u64, UINT64_C(0x4000000000000000), "%#018RX64");
2665	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xfedcba9876543210), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
2666	CHECKVAL(u64, UINT64_C(0xfedcba9876543210), "%#018RX64");
2667	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xffffffffffffffff), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
2668	CHECKVAL(u64, UINT64_C(0xffffffffffffffff), "%#018RX64");
2669	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xffffffffffffffff), UINT32_C(0xfffffff0), UINT32_C(0xffffffff));
2670	CHECKVAL(u64, UINT64_C(0xfffffff0fffffff0), "%#018RX64");
2671	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x3415934810359583), UINT32_C(0x58734981), UINT32_C(0xf8694045));
2672	CHECKVAL(u64, UINT64_C(0x128b9c3d43184763), "%#018RX64");
2673	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x3415934810359583), UINT32_C(0xf8694045), UINT32_C(0x58734981));
2674	CHECKVAL(u64, UINT64_C(0x924719355cd35a27), "%#018RX64");
2675
2676	# if 0 /* bird: question is whether this should trap or not:
2677	*
2678	* frank: Of course it must trap:
2679	*
2680	* 0xfffffff8 * 0x77d7daf8 = 0x77d7daf441412840
2681	*
2682	* During the following division, the quotient must fit into a 32-bit register.
2683	* Therefore the smallest valid divisor is
2684	*
2685	* (0x77d7daf441412840 >> 32) + 1 = 0x77d7daf5
2686	*
2687	* which is definitely greater than 0x3b9aca00.
2688	*
2689	* bird: No, the C version does not crash. So, the question is whether there's any
2690	* code depending on it not crashing.
2691	*
2692	* Of course the assembly versions of the code crash right now for the reasons you've
2693	* given, but the 32-bit MSC version does not crash.
2694	*
2695	* frank: The C version does not crash but delivers incorrect results for this case.
2696	* The reason is
2697	*
2698	* u.s.Hi = (unsigned long)(u64Hi / u32C);
2699	*
2700	* Here the division is actually 64-bit by 64-bit but the 64-bit result is truncated
2701	* to 32 bit. If using this (optimized and fast) function we should just be sure that
2702	* the operands are in a valid range.
2703	*/
2704	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xfffffff8c65d6731), UINT32_C(0x77d7daf8), UINT32_C(0x3b9aca00));
2705	CHECKVAL(u64, UINT64_C(0x02b8f9a2aa74e3dc), "%#018RX64");
2706	# endif
2707	#endif /* AMD64 \|\| X86 */
2708
2709	u32 = ASMModU64ByU32RetU32(UINT64_C(0x0ffffff8c65d6731), UINT32_C(0x77d7daf8));
2710	CHECKVAL(u32, UINT32_C(0x3B642451), "%#010RX32");
2711
2712	int32_t i32;
2713	i32 = ASMModS64ByS32RetS32(INT64_C(-11), INT32_C(-2));
2714	CHECKVAL(i32, INT32_C(-1), "%010RI32");
2715	i32 = ASMModS64ByS32RetS32(INT64_C(-11), INT32_C(2));
2716	CHECKVAL(i32, INT32_C(-1), "%010RI32");
2717	i32 = ASMModS64ByS32RetS32(INT64_C(11), INT32_C(-2));
2718	CHECKVAL(i32, INT32_C(1), "%010RI32");
2719
2720	i32 = ASMModS64ByS32RetS32(INT64_C(92233720368547758), INT32_C(2147483647));
2721	CHECKVAL(i32, INT32_C(2104533974), "%010RI32");
2722	i32 = ASMModS64ByS32RetS32(INT64_C(-92233720368547758), INT32_C(2147483647));
2723	CHECKVAL(i32, INT32_C(-2104533974), "%010RI32");
2724	}
2725
2726
2727	void tstASMByteSwap(void)
2728	{
2729	RTTestSub(g_hTest, "ASMByteSwap*");
2730
2731	uint64_t u64In = UINT64_C(0x0011223344556677);
2732	uint64_t u64Out = ASMByteSwapU64(u64In);
2733	CHECKVAL(u64In, UINT64_C(0x0011223344556677), "%#018RX64");
2734	CHECKVAL(u64Out, UINT64_C(0x7766554433221100), "%#018RX64");
2735	u64Out = ASMByteSwapU64(u64Out);
2736	CHECKVAL(u64Out, u64In, "%#018RX64");
2737	u64In = UINT64_C(0x0123456789abcdef);
2738	u64Out = ASMByteSwapU64(u64In);
2739	CHECKVAL(u64In, UINT64_C(0x0123456789abcdef), "%#018RX64");
2740	CHECKVAL(u64Out, UINT64_C(0xefcdab8967452301), "%#018RX64");
2741	u64Out = ASMByteSwapU64(u64Out);
2742	CHECKVAL(u64Out, u64In, "%#018RX64");
2743	u64In = 0;
2744	u64Out = ASMByteSwapU64(u64In);
2745	CHECKVAL(u64Out, u64In, "%#018RX64");
2746	u64In = UINT64_MAX;
2747	u64Out = ASMByteSwapU64(u64In);
2748	CHECKVAL(u64Out, u64In, "%#018RX64");
2749
2750	uint32_t u32In = UINT32_C(0x00112233);
2751	uint32_t u32Out = ASMByteSwapU32(u32In);
2752	CHECKVAL(u32In, UINT32_C(0x00112233), "%#010RX32");
2753	CHECKVAL(u32Out, UINT32_C(0x33221100), "%#010RX32");
2754	u32Out = ASMByteSwapU32(u32Out);
2755	CHECKVAL(u32Out, u32In, "%#010RX32");
2756	u32In = UINT32_C(0x12345678);
2757	u32Out = ASMByteSwapU32(u32In);
2758	CHECKVAL(u32In, UINT32_C(0x12345678), "%#010RX32");
2759	CHECKVAL(u32Out, UINT32_C(0x78563412), "%#010RX32");
2760	u32Out = ASMByteSwapU32(u32Out);
2761	CHECKVAL(u32Out, u32In, "%#010RX32");
2762	u32In = 0;
2763	u32Out = ASMByteSwapU32(u32In);
2764	CHECKVAL(u32Out, u32In, "%#010RX32");
2765	u32In = UINT32_MAX;
2766	u32Out = ASMByteSwapU32(u32In);
2767	CHECKVAL(u32Out, u32In, "%#010RX32");
2768
2769	uint16_t u16In = UINT16_C(0x0011);
2770	uint16_t u16Out = ASMByteSwapU16(u16In);
2771	CHECKVAL(u16In, UINT16_C(0x0011), "%#06RX16");
2772	CHECKVAL(u16Out, UINT16_C(0x1100), "%#06RX16");
2773	u16Out = ASMByteSwapU16(u16Out);
2774	CHECKVAL(u16Out, u16In, "%#06RX16");
2775	u16In = UINT16_C(0x1234);
2776	u16Out = ASMByteSwapU16(u16In);
2777	CHECKVAL(u16In, UINT16_C(0x1234), "%#06RX16");
2778	CHECKVAL(u16Out, UINT16_C(0x3412), "%#06RX16");
2779	u16Out = ASMByteSwapU16(u16Out);
2780	CHECKVAL(u16Out, u16In, "%#06RX16");
2781	u16In = 0;
2782	u16Out = ASMByteSwapU16(u16In);
2783	CHECKVAL(u16Out, u16In, "%#06RX16");
2784	u16In = UINT16_MAX;
2785	u16Out = ASMByteSwapU16(u16In);
2786	CHECKVAL(u16Out, u16In, "%#06RX16");
2787	}
2788
2789
2790	void tstASMBench(void)
2791	{
2792	/*
2793	* Make this static. We don't want to have this located on the stack.
2794	*/
2795	static uint8_t volatile s_u8;
2796	static int8_t volatile s_i8;
2797	static uint16_t volatile s_u16;
2798	static int16_t volatile s_i16;
2799	static uint32_t volatile s_u32;
2800	static int32_t volatile s_i32;
2801	static uint64_t volatile s_u64;
2802	static int64_t volatile s_i64;
2803	static uint8_t s_u8Old;
2804	static int8_t s_i8Old;
2805	static uint16_t s_u16Old;
2806	static int16_t s_i16Old;
2807	static uint32_t s_u32Old;
2808	static int32_t s_i32Old;
2809	static uint64_t s_u64Old;
2810	static int64_t s_i64Old;
2811	unsigned i;
2812	const unsigned cRounds = _16M; /* Must be multiple of 8 */
2813	uint64_t u64Elapsed;
2814
2815	RTTestSub(g_hTest, "Benchmarking");
2816
2817	#if 0 && !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86) \|\| defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32))
2818	# define BENCH(op, str) \
2819	do { \
2820	RTThreadYield(); \
2821	u64Elapsed = ASMReadTSC(); \
2822	for (i = cRounds; i > 0; i--) \
2823	op; \
2824	u64Elapsed = ASMReadTSC() - u64Elapsed; \
2825	RTTestValue(g_hTest, str, u64Elapsed / cRounds, RTTESTUNIT_TICKS_PER_CALL); \
2826	} while (0)
2827	#else
2828	# define BENCH(op, str) \
2829	do { \
2830	RTThreadYield(); \
2831	u64Elapsed = RTTimeNanoTS(); \
2832	for (i = cRounds / 8; i > 0; i--) \
2833	{ \
2834	op; \
2835	op; \
2836	op; \
2837	op; \
2838	op; \
2839	op; \
2840	op; \
2841	op; \
2842	} \
2843	u64Elapsed = RTTimeNanoTS() - u64Elapsed; \
2844	RTTestValue(g_hTest, str, u64Elapsed * 1000 / cRounds, RTTESTUNIT_PS_PER_CALL); \
2845	} while (0)
2846	#endif
2847	#if (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86) \|\| defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32)) && !defined(GCC44_32BIT_PIC)
2848	# define BENCH_TSC(op, str) \
2849	do { \
2850	RTThreadYield(); \
2851	u64Elapsed = ASMReadTSC(); \
2852	for (i = cRounds / 8; i > 0; i--) \
2853	{ \
2854	op; \
2855	op; \
2856	op; \
2857	op; \
2858	op; \
2859	op; \
2860	op; \
2861	op; \
2862	} \
2863	u64Elapsed = ASMReadTSC() - u64Elapsed; \
2864	RTTestValue(g_hTest, str, u64Elapsed / cRounds, RTTESTUNIT_TICKS_PER_CALL); \
2865	} while (0)
2866	#else
2867	# define BENCH_TSC(op, str) BENCH(op, str)
2868	#endif
2869
2870	BENCH(s_u32 = 0, "s_u32 = 0");
2871	BENCH(ASMAtomicUoReadU8(&s_u8), "ASMAtomicUoReadU8");
2872	BENCH(ASMAtomicUoReadS8(&s_i8), "ASMAtomicUoReadS8");
2873	BENCH(ASMAtomicUoReadU16(&s_u16), "ASMAtomicUoReadU16");
2874	BENCH(ASMAtomicUoReadS16(&s_i16), "ASMAtomicUoReadS16");
2875	BENCH(ASMAtomicUoReadU32(&s_u32), "ASMAtomicUoReadU32");
2876	BENCH(ASMAtomicUoReadS32(&s_i32), "ASMAtomicUoReadS32");
2877	BENCH(ASMAtomicUoReadU64(&s_u64), "ASMAtomicUoReadU64");
2878	BENCH(ASMAtomicUoReadS64(&s_i64), "ASMAtomicUoReadS64");
2879	BENCH(ASMAtomicReadU8(&s_u8), "ASMAtomicReadU8");
2880	BENCH(ASMAtomicReadS8(&s_i8), "ASMAtomicReadS8");
2881	BENCH(ASMAtomicReadU16(&s_u16), "ASMAtomicReadU16");
2882	BENCH(ASMAtomicReadS16(&s_i16), "ASMAtomicReadS16");
2883	BENCH(ASMAtomicReadU32(&s_u32), "ASMAtomicReadU32");
2884	BENCH(ASMAtomicReadS32(&s_i32), "ASMAtomicReadS32");
2885	BENCH(ASMAtomicReadU64(&s_u64), "ASMAtomicReadU64");
2886	BENCH(ASMAtomicReadS64(&s_i64), "ASMAtomicReadS64");
2887	BENCH(ASMAtomicUoWriteU8(&s_u8, 0), "ASMAtomicUoWriteU8");
2888	BENCH(ASMAtomicUoWriteS8(&s_i8, 0), "ASMAtomicUoWriteS8");
2889	BENCH(ASMAtomicUoWriteU16(&s_u16, 0), "ASMAtomicUoWriteU16");
2890	BENCH(ASMAtomicUoWriteS16(&s_i16, 0), "ASMAtomicUoWriteS16");
2891	BENCH(ASMAtomicUoWriteU32(&s_u32, 0), "ASMAtomicUoWriteU32");
2892	BENCH(ASMAtomicUoWriteS32(&s_i32, 0), "ASMAtomicUoWriteS32");
2893	BENCH(ASMAtomicUoWriteU64(&s_u64, 0), "ASMAtomicUoWriteU64");
2894	BENCH(ASMAtomicUoWriteS64(&s_i64, 0), "ASMAtomicUoWriteS64");
2895	BENCH(ASMAtomicWriteU8(&s_u8, 0), "ASMAtomicWriteU8");
2896	BENCH(ASMAtomicWriteS8(&s_i8, 0), "ASMAtomicWriteS8");
2897	BENCH(ASMAtomicWriteU16(&s_u16, 0), "ASMAtomicWriteU16");
2898	BENCH(ASMAtomicWriteS16(&s_i16, 0), "ASMAtomicWriteS16");
2899	BENCH(ASMAtomicWriteU32(&s_u32, 0), "ASMAtomicWriteU32");
2900	BENCH(ASMAtomicWriteS32(&s_i32, 0), "ASMAtomicWriteS32");
2901	BENCH(ASMAtomicWriteU64(&s_u64, 0), "ASMAtomicWriteU64");
2902	BENCH(ASMAtomicWriteS64(&s_i64, 0), "ASMAtomicWriteS64");
2903	BENCH(ASMAtomicXchgU8(&s_u8, 0), "ASMAtomicXchgU8");
2904	BENCH(ASMAtomicXchgS8(&s_i8, 0), "ASMAtomicXchgS8");
2905	BENCH(ASMAtomicXchgU16(&s_u16, 0), "ASMAtomicXchgU16");
2906	BENCH(ASMAtomicXchgS16(&s_i16, 0), "ASMAtomicXchgS16");
2907	BENCH(ASMAtomicXchgU32(&s_u32, 0), "ASMAtomicXchgU32");
2908	BENCH(ASMAtomicXchgS32(&s_i32, 0), "ASMAtomicXchgS32");
2909	BENCH(ASMAtomicXchgU64(&s_u64, 0), "ASMAtomicXchgU64");
2910	BENCH(ASMAtomicXchgS64(&s_i64, 0), "ASMAtomicXchgS64");
2911	BENCH(ASMAtomicCmpXchgU8(&s_u8, 0, 0), "ASMAtomicCmpXchgU8");
2912	BENCH(ASMAtomicCmpXchgS8(&s_i8, 0, 0), "ASMAtomicCmpXchgS8");
2913	//BENCH(ASMAtomicCmpXchgU16(&s_u16, 0, 0), "ASMAtomicCmpXchgU16");
2914	//BENCH(ASMAtomicCmpXchgS16(&s_i16, 0, 0), "ASMAtomicCmpXchgS16");
2915	BENCH(ASMAtomicCmpXchgU32(&s_u32, 0, 0), "ASMAtomicCmpXchgU32");
2916	BENCH(ASMAtomicCmpXchgS32(&s_i32, 0, 0), "ASMAtomicCmpXchgS32");
2917	BENCH(ASMAtomicCmpXchgU64(&s_u64, 0, 0), "ASMAtomicCmpXchgU64");
2918	BENCH(ASMAtomicCmpXchgS64(&s_i64, 0, 0), "ASMAtomicCmpXchgS64");
2919	BENCH(ASMAtomicCmpXchgU8(&s_u8, 0, 1), "ASMAtomicCmpXchgU8/neg");
2920	BENCH(ASMAtomicCmpXchgS8(&s_i8, 0, 1), "ASMAtomicCmpXchgS8/neg");
2921	//BENCH(ASMAtomicCmpXchgU16(&s_u16, 0, 1), "ASMAtomicCmpXchgU16/neg");
2922	//BENCH(ASMAtomicCmpXchgS16(&s_s16, 0, 1), "ASMAtomicCmpXchgS16/neg");
2923	BENCH(ASMAtomicCmpXchgU32(&s_u32, 0, 1), "ASMAtomicCmpXchgU32/neg");
2924	BENCH(ASMAtomicCmpXchgS32(&s_i32, 0, 1), "ASMAtomicCmpXchgS32/neg");
2925	BENCH(ASMAtomicCmpXchgU64(&s_u64, 0, 1), "ASMAtomicCmpXchgU64/neg");
2926	BENCH(ASMAtomicCmpXchgS64(&s_i64, 0, 1), "ASMAtomicCmpXchgS64/neg");
2927	BENCH(ASMAtomicCmpXchgExU8(&s_u8, 0, 0, &s_u8Old), "ASMAtomicCmpXchgExU8");
2928	BENCH(ASMAtomicCmpXchgExS8(&s_i8, 0, 0, &s_i8Old), "ASMAtomicCmpXchgExS8");
2929	BENCH(ASMAtomicCmpXchgExU16(&s_u16, 0, 0, &s_u16Old), "ASMAtomicCmpXchgExU16");
2930	BENCH(ASMAtomicCmpXchgExS16(&s_i16, 0, 0, &s_i16Old), "ASMAtomicCmpXchgExS16");
2931	BENCH(ASMAtomicCmpXchgExU32(&s_u32, 0, 0, &s_u32Old), "ASMAtomicCmpXchgExU32");
2932	BENCH(ASMAtomicCmpXchgExS32(&s_i32, 0, 0, &s_i32Old), "ASMAtomicCmpXchgExS32");
2933	BENCH(ASMAtomicCmpXchgExU64(&s_u64, 0, 0, &s_u64Old), "ASMAtomicCmpXchgExU64");
2934	BENCH(ASMAtomicCmpXchgExS64(&s_i64, 0, 0, &s_i64Old), "ASMAtomicCmpXchgExS64");
2935	BENCH(ASMAtomicCmpXchgExU8(&s_u8, 0, 1, &s_u8Old), "ASMAtomicCmpXchgExU8/neg");
2936	BENCH(ASMAtomicCmpXchgExS8(&s_i8, 0, 1, &s_i8Old), "ASMAtomicCmpXchgExS8/neg");
2937	BENCH(ASMAtomicCmpXchgExU16(&s_u16, 0, 1, &s_u16Old), "ASMAtomicCmpXchgExU16/neg");
2938	BENCH(ASMAtomicCmpXchgExS16(&s_i16, 0, 1, &s_i16Old), "ASMAtomicCmpXchgExS16/neg");
2939	BENCH(ASMAtomicCmpXchgExU32(&s_u32, 0, 1, &s_u32Old), "ASMAtomicCmpXchgExU32/neg");
2940	BENCH(ASMAtomicCmpXchgExS32(&s_i32, 0, 1, &s_i32Old), "ASMAtomicCmpXchgExS32/neg");
2941	BENCH(ASMAtomicCmpXchgExU64(&s_u64, 0, 1, &s_u64Old), "ASMAtomicCmpXchgExU64/neg");
2942	BENCH(ASMAtomicCmpXchgExS64(&s_i64, 0, 1, &s_i64Old), "ASMAtomicCmpXchgExS64/neg");
2943	BENCH(ASMAtomicIncU32(&s_u32), "ASMAtomicIncU32");
2944	BENCH(ASMAtomicIncS32(&s_i32), "ASMAtomicIncS32");
2945	BENCH(ASMAtomicDecU32(&s_u32), "ASMAtomicDecU32");
2946	BENCH(ASMAtomicDecS32(&s_i32), "ASMAtomicDecS32");
2947	BENCH(ASMAtomicAddU32(&s_u32, 5), "ASMAtomicAddU32");
2948	BENCH(ASMAtomicAddS32(&s_i32, 5), "ASMAtomicAddS32");
2949	BENCH(ASMAtomicUoIncU32(&s_u32), "ASMAtomicUoIncU32");
2950	BENCH(ASMAtomicUoDecU32(&s_u32), "ASMAtomicUoDecU32");
2951	BENCH(ASMAtomicUoAndU32(&s_u32, 0xffffffff), "ASMAtomicUoAndU32");
2952	BENCH(ASMAtomicUoOrU32(&s_u32, 0xffffffff), "ASMAtomicUoOrU32");
2953	#if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
2954	BENCH_TSC(ASMSerializeInstructionCpuId(), "ASMSerializeInstructionCpuId");
2955	BENCH_TSC(ASMSerializeInstructionIRet(), "ASMSerializeInstructionIRet");
2956	#endif
2957	BENCH(ASMReadFence(), "ASMReadFence");
2958	BENCH(ASMWriteFence(), "ASMWriteFence");
2959	BENCH(ASMMemoryFence(), "ASMMemoryFence");
2960	BENCH(ASMSerializeInstruction(), "ASMSerializeInstruction");
2961	BENCH(ASMNopPause(), "ASMNopPause");
2962
2963	/* The Darwin gcc does not like this ... */
2964	#if !defined(RT_OS_DARWIN) && !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
2965	BENCH(s_u8 = ASMGetApicId(), "ASMGetApicId");
2966	BENCH(s_u32 = ASMGetApicIdExt0B(), "ASMGetApicIdExt0B");
2967	BENCH(s_u32 = ASMGetApicIdExt8000001E(), "ASMGetApicIdExt8000001E");
2968	#endif
2969	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86) \|\| defined(RT_ARCH_ARM64) \|\| defined(RT_ARCH_ARM32))
2970	BENCH(s_u64 = ASMReadTSC(), "ASMReadTSC");
2971	#endif
2972	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
2973	uint32_t uAux;
2974	if ( ASMHasCpuId()
2975	&& RTX86IsValidExtRange(ASMCpuId_EAX(0x80000000))
2976	&& (ASMCpuId_EDX(0x80000001) & X86_CPUID_EXT_FEATURE_EDX_RDTSCP) )
2977	{
2978	BENCH_TSC(ASMSerializeInstructionRdTscp(), "ASMSerializeInstructionRdTscp");
2979	BENCH(s_u64 = ASMReadTscWithAux(&uAux), "ASMReadTscWithAux");
2980	}
2981	union
2982	{
2983	uint64_t u64[2];
2984	RTIDTR Unaligned;
2985	struct
2986	{
2987	uint16_t abPadding[3];
2988	RTIDTR Aligned;
2989	} s;
2990	} uBuf;
2991	Assert(((uintptr_t)&uBuf.Unaligned.pIdt & (sizeof(uintptr_t) - 1)) != 0);
2992	BENCH(ASMGetIDTR(&uBuf.Unaligned), "ASMGetIDTR/unaligned");
2993	Assert(((uintptr_t)&uBuf.s.Aligned.pIdt & (sizeof(uintptr_t) - 1)) == 0);
2994	BENCH(ASMGetIDTR(&uBuf.s.Aligned), "ASMGetIDTR/aligned");
2995	#endif
2996
2997	#undef BENCH
2998	}
2999
3000
3001	int main(int argc, char **argv)
3002	{
3003	RT_NOREF_PV(argc); RT_NOREF_PV(argv);
3004
3005	int rc = RTTestInitAndCreate("tstRTInlineAsm", &g_hTest);
3006	if (rc)
3007	return rc;
3008	RTTestBanner(g_hTest);
3009
3010	/*
3011	* Execute the tests.
3012	*/
3013	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
3014	tstASMCpuId();
3015	//bruteForceCpuId();
3016	#endif
3017	#if 1
3018	tstASMAtomicRead();
3019	tstASMAtomicWrite();
3020	tstASMAtomicXchg();
3021	tstASMAtomicCmpXchg();
3022	tstASMAtomicCmpXchgEx();
3023
3024	tstASMAtomicAdd();
3025	tstASMAtomicDecInc();
3026	tstASMAtomicAndOrXor();
3027
3028	tstASMMemZeroPage();
3029	tstASMMemIsZeroPage(g_hTest);
3030	tstASMMemFirstMismatchingU8(g_hTest);
3031	tstASMMemZero32();
3032	tstASMMemFill32();
3033	tstASMProbe(g_hTest);
3034
3035	tstASMMisc();
3036
3037	tstASMMath();
3038
3039	tstASMByteSwap();
3040
3041	tstASMBench();
3042	#endif
3043
3044	/*
3045	* Show the result.
3046	*/
3047	return RTTestSummaryAndDestroy(g_hTest);
3048	}
3049

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

source: vbox/trunk/src/VBox/Runtime/testcase/tstRTInlineAsm.cpp@ 93755

Download in other formats: