tstRTInlineAsm.cpp@ 85546

Last change on this file since 85546 was 85072, checked in by vboxsync, 4 years ago
tstRTInlineAsm.cpp: Stop using the 'register' keyword, Clang barfs on it in C++11 mode. bugref:9790
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 76.1 KB

Line
1	/* $Id: tstRTInlineAsm.cpp 85072 2020-07-06 14:42:28Z vboxsync $ */
2	/** @file
3	* IPRT Testcase - inline assembly.
4	*/
5
6	/*
7	* Copyright (C) 2006-2020 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	#else
47	# include <iprt/time.h>
48	#endif
49	#include <iprt/rand.h>
50	#include <iprt/stream.h>
51	#include <iprt/string.h>
52	#include <iprt/param.h>
53	#include <iprt/thread.h>
54	#include <iprt/test.h>
55	#include <iprt/time.h>
56
57
58
59	/*********************************************************************************************************************************
60	* Defined Constants And Macros *
61	*********************************************************************************************************************************/
62	#define CHECKVAL(val, expect, fmt) \
63	do \
64	{ \
65	if ((val) != (expect)) \
66	{ \
67	RTTestFailed(g_hTest, "%s, %d: " #val ": expected " fmt " got " fmt "\n", __FUNCTION__, __LINE__, (expect), (val)); \
68	} \
69	} while (0)
70
71	#define CHECKOP(op, expect, fmt, type) \
72	do \
73	{ \
74	type val = op; \
75	if (val != (type)(expect)) \
76	{ \
77	RTTestFailed(g_hTest, "%s, %d: " #op ": expected " fmt " got " fmt "\n", __FUNCTION__, __LINE__, (type)(expect), val); \
78	} \
79	} while (0)
80
81	/**
82	* Calls a worker function with different worker variable storage types.
83	*/
84	#define DO_SIMPLE_TEST(name, type) \
85	do \
86	{ \
87	RTTestISub(#name); \
88	type StackVar; \
89	tst ## name ## Worker(&StackVar); \
90	\
91	type pVar = (type )RTTestGuardedAllocHead(g_hTest, sizeof(type)); \
92	RTTEST_CHECK_BREAK(g_hTest, pVar); \
93	tst ## name ## Worker(pVar); \
94	RTTestGuardedFree(g_hTest, pVar); \
95	\
96	pVar = (type *)RTTestGuardedAllocTail(g_hTest, sizeof(type)); \
97	RTTEST_CHECK_BREAK(g_hTest, pVar); \
98	tst ## name ## Worker(pVar); \
99	RTTestGuardedFree(g_hTest, pVar); \
100	} while (0)
101
102
103	/*********************************************************************************************************************************
104	* Global Variables *
105	*********************************************************************************************************************************/
106	/** The test instance. */
107	static RTTEST g_hTest;
108
109
110
111	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
112
113	const char *getCacheAss(unsigned u)
114	{
115	if (u == 0)
116	return "res0 ";
117	if (u == 1)
118	return "direct";
119	if (u >= 256)
120	return "???";
121
122	char *pszRet;
123	RTStrAPrintf(&pszRet, "%d way", u); /* intentional leak! */
124	return pszRet;
125	}
126
127
128	const char *getL2CacheAss(unsigned u)
129	{
130	switch (u)
131	{
132	case 0: return "off ";
133	case 1: return "direct";
134	case 2: return "2 way ";
135	case 3: return "res3 ";
136	case 4: return "4 way ";
137	case 5: return "res5 ";
138	case 6: return "8 way ";
139	case 7: return "res7 ";
140	case 8: return "16 way";
141	case 9: return "res9 ";
142	case 10: return "res10 ";
143	case 11: return "res11 ";
144	case 12: return "res12 ";
145	case 13: return "res13 ";
146	case 14: return "res14 ";
147	case 15: return "fully ";
148	default:
149	return "????";
150	}
151	}
152
153
154	/**
155	* Test and dump all possible info from the CPUID instruction.
156	*
157	* @remark Bits shared with the libc cpuid.c program. This all written by me, so no worries.
158	* @todo transform the dumping into a generic runtime function. We'll need it for logging!
159	*/
160	void tstASMCpuId(void)
161	{
162	RTTestISub("ASMCpuId");
163
164	unsigned iBit;
165	struct
166	{
167	uint32_t uEBX, uEAX, uEDX, uECX;
168	} s;
169	if (!ASMHasCpuId())
170	{
171	RTTestIPrintf(RTTESTLVL_ALWAYS, "warning! CPU doesn't support CPUID\n");
172	return;
173	}
174
175	/*
176	* Try the 0 function and use that for checking the ASMCpuId_* variants.
177	*/
178	ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
179
180	uint32_t u32;
181
182	u32 = ASMCpuId_EAX(0);
183	CHECKVAL(u32, s.uEAX, "%x");
184	u32 = ASMCpuId_EBX(0);
185	CHECKVAL(u32, s.uEBX, "%x");
186	u32 = ASMCpuId_ECX(0);
187	CHECKVAL(u32, s.uECX, "%x");
188	u32 = ASMCpuId_EDX(0);
189	CHECKVAL(u32, s.uEDX, "%x");
190
191	uint32_t uECX2 = s.uECX - 1;
192	uint32_t uEDX2 = s.uEDX - 1;
193	ASMCpuId_ECX_EDX(0, &uECX2, &uEDX2);
194	CHECKVAL(uECX2, s.uECX, "%x");
195	CHECKVAL(uEDX2, s.uEDX, "%x");
196
197	uint32_t uEAX2 = s.uEAX - 1;
198	uint32_t uEBX2 = s.uEBX - 1;
199	uECX2 = s.uECX - 1;
200	uEDX2 = s.uEDX - 1;
201	ASMCpuIdExSlow(0, 0, 0, 0, &uEAX2, &uEBX2, &uECX2, &uEDX2);
202	CHECKVAL(uEAX2, s.uEAX, "%x");
203	CHECKVAL(uEBX2, s.uEBX, "%x");
204	CHECKVAL(uECX2, s.uECX, "%x");
205	CHECKVAL(uEDX2, s.uEDX, "%x");
206
207	/*
208	* Check the extended APIC stuff.
209	*/
210	uint32_t idExtApic;
211	if (ASMCpuId_EAX(0) >= 0xb)
212	{
213	uint8_t idApic = ASMGetApicId();
214	do
215	{
216	uEAX2 = uEBX2 = uECX2 = uEDX2 = UINT32_C(0x50486744);
217	ASMCpuIdExSlow(0xb, 0, 0, 0, &uEAX2, &uEBX2, &uECX2, &uEDX2);
218	idExtApic = ASMGetApicIdExt0B();
219	} while (ASMGetApicId() != idApic);
220
221	CHECKVAL(uEDX2, idExtApic, "%x");
222	if (idApic != (uint8_t)idExtApic && uECX2 != 0)
223	RTTestIFailed("ASMGetApicIdExt0B() -> %#x vs ASMGetApicId() -> %#x", idExtApic, idApic);
224	}
225	if (ASMCpuId_EAX(UINT32_C(0x80000000)) >= UINT32_C(0x8000001E))
226	{
227	uint8_t idApic = ASMGetApicId();
228	do
229	{
230	uEAX2 = uEBX2 = uECX2 = uEDX2 = UINT32_C(0x50486744);
231	ASMCpuIdExSlow(0x8000001e, 0, 0, 0, &uEAX2, &uEBX2, &uECX2, &uEDX2);
232	idExtApic = ASMGetApicIdExt8000001E();
233	} while (ASMGetApicId() != idApic);
234	CHECKVAL(uEAX2, idExtApic, "%x");
235	if (idApic != (uint8_t)idExtApic)
236	RTTestIFailed("ASMGetApicIdExt8000001E() -> %#x vs ASMGetApicId() -> %#x", idExtApic, idApic);
237	}
238
239	/*
240	* Done testing, dump the information.
241	*/
242	RTTestIPrintf(RTTESTLVL_ALWAYS, "CPUID Dump\n");
243	ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
244	const uint32_t cFunctions = s.uEAX;
245
246	/* raw dump */
247	RTTestIPrintf(RTTESTLVL_ALWAYS,
248	"\n"
249	" RAW Standard CPUIDs\n"
250	"Function eax ebx ecx edx\n");
251	for (unsigned iStd = 0; iStd <= cFunctions + 3; iStd++)
252	{
253	ASMCpuId_Idx_ECX(iStd, 0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
254	RTTestIPrintf(RTTESTLVL_ALWAYS, "%08x %08x %08x %08x %08x%s\n",
255	iStd, s.uEAX, s.uEBX, s.uECX, s.uEDX, iStd <= cFunctions ? "" : "*");
256
257	/* Some leafs output depend on the initial value of ECX.
258	* The same seems to apply to invalid standard functions */
259	if (iStd > cFunctions)
260	continue;
261	if (iStd == 0x04) /* Deterministic Cache Parameters Leaf */
262	for (uint32_t uECX = 1; s.uEAX & 0x1f; uECX++)
263	{
264	ASMCpuId_Idx_ECX(iStd, uECX, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
265	RTTestIPrintf(RTTESTLVL_ALWAYS, " [%02x] %08x %08x %08x %08x\n", uECX, s.uEAX, s.uEBX, s.uECX, s.uEDX);
266	RTTESTI_CHECK_BREAK(uECX < 128);
267	}
268	else if (iStd == 0x07) /* Structured Extended Feature Flags */
269	{
270	uint32_t uMax = s.uEAX;
271	for (uint32_t uECX = 1; uECX < uMax; uECX++)
272	{
273	ASMCpuId_Idx_ECX(iStd, uECX, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
274	RTTestIPrintf(RTTESTLVL_ALWAYS, " [%02x] %08x %08x %08x %08x\n", uECX, s.uEAX, s.uEBX, s.uECX, s.uEDX);
275	RTTESTI_CHECK_BREAK(uECX < 128);
276	}
277	}
278	else if (iStd == 0x0b) /* Extended Topology Enumeration Leafs */
279	for (uint32_t uECX = 1; (s.uEAX & 0x1f) && (s.uEBX & 0xffff); uECX++)
280	{
281	ASMCpuId_Idx_ECX(iStd, uECX, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
282	RTTestIPrintf(RTTESTLVL_ALWAYS, " [%02x] %08x %08x %08x %08x\n", uECX, s.uEAX, s.uEBX, s.uECX, s.uEDX);
283	RTTESTI_CHECK_BREAK(uECX < 128);
284	}
285	else if (iStd == 0x0d) /* Extended State Enumeration Leafs */
286	for (uint32_t uECX = 1; s.uEAX != 0 \|\| s.uEBX != 0 \|\| s.uECX != 0 \|\| s.uEDX != 0; uECX++)
287	{
288	ASMCpuId_Idx_ECX(iStd, uECX, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
289	RTTestIPrintf(RTTESTLVL_ALWAYS, " [%02x] %08x %08x %08x %08x\n", uECX, s.uEAX, s.uEBX, s.uECX, s.uEDX);
290	RTTESTI_CHECK_BREAK(uECX < 128);
291	}
292	else if ( iStd == 0x0f /* Platform quality of service monitoring (PQM) */
293	\|\| iStd == 0x10 /* Platform quality of service enforcement (PQE) */
294	\|\| iStd == 0x12 /* SGX Enumeration */
295	\|\| iStd == 0x14 /* Processor Trace Enumeration */
296	\|\| iStd == 0x17 /* SoC Vendor Attribute Enumeration */
297	\|\| iStd == 0x18 /* Deterministic Address Translation Parameters */)
298	{
299	/** @todo */
300	}
301	else
302	{
303	u32 = ASMCpuId_EAX(iStd);
304	CHECKVAL(u32, s.uEAX, "%x");
305
306	uint32_t u32EbxMask = UINT32_MAX;
307	if (iStd == 1)
308	u32EbxMask = UINT32_C(0x00ffffff); /* Omit the local apic ID in case we're rescheduled. */
309	u32 = ASMCpuId_EBX(iStd);
310	CHECKVAL(u32 & u32EbxMask, s.uEBX & u32EbxMask, "%x");
311
312	u32 = ASMCpuId_ECX(iStd);
313	CHECKVAL(u32, s.uECX, "%x");
314	u32 = ASMCpuId_EDX(iStd);
315	CHECKVAL(u32, s.uEDX, "%x");
316
317	uECX2 = s.uECX - 1;
318	uEDX2 = s.uEDX - 1;
319	ASMCpuId_ECX_EDX(iStd, &uECX2, &uEDX2);
320	CHECKVAL(uECX2, s.uECX, "%x");
321	CHECKVAL(uEDX2, s.uEDX, "%x");
322
323	uEAX2 = s.uEAX - 1;
324	uEBX2 = s.uEBX - 1;
325	uECX2 = s.uECX - 1;
326	uEDX2 = s.uEDX - 1;
327	ASMCpuId(iStd, &uEAX2, &uEBX2, &uECX2, &uEDX2);
328	CHECKVAL(uEAX2, s.uEAX, "%x");
329	CHECKVAL(uEBX2 & u32EbxMask, s.uEBX & u32EbxMask, "%x");
330	CHECKVAL(uECX2, s.uECX, "%x");
331	CHECKVAL(uEDX2, s.uEDX, "%x");
332	}
333	}
334
335	/*
336	* Understandable output
337	*/
338	ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
339	RTTestIPrintf(RTTESTLVL_ALWAYS,
340	"Name: %.04s%.04s%.04s\n"
341	"Support: 0-%u\n",
342	&s.uEBX, &s.uEDX, &s.uECX, s.uEAX);
343	bool const fIntel = ASMIsIntelCpuEx(s.uEBX, s.uECX, s.uEDX);
344
345	/*
346	* Get Features.
347	*/
348	if (cFunctions >= 1)
349	{
350	static const char * const s_apszTypes[4] = { "primary", "overdrive", "MP", "reserved" };
351	ASMCpuId(1, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
352	RTTestIPrintf(RTTESTLVL_ALWAYS,
353	"Family: %#x \tExtended: %#x \tEffective: %#x\n"
354	"Model: %#x \tExtended: %#x \tEffective: %#x\n"
355	"Stepping: %d\n"
356	"Type: %d (%s)\n"
357	"APIC ID: %#04x\n"
358	"Logical CPUs: %d\n"
359	"CLFLUSH Size: %d\n"
360	"Brand ID: %#04x\n",
361	(s.uEAX >> 8) & 0xf, (s.uEAX >> 20) & 0x7f, ASMGetCpuFamily(s.uEAX),
362	(s.uEAX >> 4) & 0xf, (s.uEAX >> 16) & 0x0f, ASMGetCpuModel(s.uEAX, fIntel),
363	ASMGetCpuStepping(s.uEAX),
364	(s.uEAX >> 12) & 0x3, s_apszTypes[(s.uEAX >> 12) & 0x3],
365	(s.uEBX >> 24) & 0xff,
366	(s.uEBX >> 16) & 0xff,
367	(s.uEBX >> 8) & 0xff,
368	(s.uEBX >> 0) & 0xff);
369
370	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features EDX: ");
371	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FPU");
372	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VME");
373	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DE");
374	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE");
375	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TSC");
376	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MSR");
377	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAE");
378	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCE");
379	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CX8");
380	if (s.uEDX & RT_BIT(9)) RTTestIPrintf(RTTESTLVL_ALWAYS, " APIC");
381	if (s.uEDX & RT_BIT(10)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 10");
382	if (s.uEDX & RT_BIT(11)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SEP");
383	if (s.uEDX & RT_BIT(12)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MTRR");
384	if (s.uEDX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PGE");
385	if (s.uEDX & RT_BIT(14)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCA");
386	if (s.uEDX & RT_BIT(15)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMOV");
387	if (s.uEDX & RT_BIT(16)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAT");
388	if (s.uEDX & RT_BIT(17)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE36");
389	if (s.uEDX & RT_BIT(18)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSN");
390	if (s.uEDX & RT_BIT(19)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CLFSH");
391	if (s.uEDX & RT_BIT(20)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 20");
392	if (s.uEDX & RT_BIT(21)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DS");
393	if (s.uEDX & RT_BIT(22)) RTTestIPrintf(RTTESTLVL_ALWAYS, " ACPI");
394	if (s.uEDX & RT_BIT(23)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MMX");
395	if (s.uEDX & RT_BIT(24)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FXSR");
396	if (s.uEDX & RT_BIT(25)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE");
397	if (s.uEDX & RT_BIT(26)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE2");
398	if (s.uEDX & RT_BIT(27)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SS");
399	if (s.uEDX & RT_BIT(28)) RTTestIPrintf(RTTESTLVL_ALWAYS, " HTT");
400	if (s.uEDX & RT_BIT(29)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 29");
401	if (s.uEDX & RT_BIT(30)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 30");
402	if (s.uEDX & RT_BIT(31)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 31");
403	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
404
405	/** @todo check intel docs. */
406	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features ECX: ");
407	if (s.uECX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE3");
408	for (iBit = 1; iBit < 13; iBit++)
409	if (s.uECX & RT_BIT(iBit))
410	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
411	if (s.uECX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CX16");
412	for (iBit = 14; iBit < 32; iBit++)
413	if (s.uECX & RT_BIT(iBit))
414	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
415	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
416	}
417
418	/*
419	* Extended.
420	* Implemented after AMD specs.
421	*/
422	/** @todo check out the intel specs. */
423	ASMCpuId(0x80000000, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
424	if (!s.uEAX && !s.uEBX && !s.uECX && !s.uEDX)
425	{
426	RTTestIPrintf(RTTESTLVL_ALWAYS, "No extended CPUID info? Check the manual on how to detect this...\n");
427	return;
428	}
429	const uint32_t cExtFunctions = s.uEAX \| 0x80000000;
430
431	/* raw dump */
432	RTTestIPrintf(RTTESTLVL_ALWAYS,
433	"\n"
434	" RAW Extended CPUIDs\n"
435	"Function eax ebx ecx edx\n");
436	for (unsigned iExt = 0x80000000; iExt <= cExtFunctions + 3; iExt++)
437	{
438	ASMCpuId(iExt, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
439	RTTestIPrintf(RTTESTLVL_ALWAYS, "%08x %08x %08x %08x %08x%s\n",
440	iExt, s.uEAX, s.uEBX, s.uECX, s.uEDX, iExt <= cExtFunctions ? "" : "*");
441
442	if (iExt > cExtFunctions)
443	continue; /* Invalid extended functions seems change the value if ECX changes */
444	if (iExt == 0x8000001d)
445	continue; /* Takes cache level in ecx. */
446
447	u32 = ASMCpuId_EAX(iExt);
448	CHECKVAL(u32, s.uEAX, "%x");
449	u32 = ASMCpuId_EBX(iExt);
450	CHECKVAL(u32, s.uEBX, "%x");
451	u32 = ASMCpuId_ECX(iExt);
452	CHECKVAL(u32, s.uECX, "%x");
453	u32 = ASMCpuId_EDX(iExt);
454	CHECKVAL(u32, s.uEDX, "%x");
455
456	uECX2 = s.uECX - 1;
457	uEDX2 = s.uEDX - 1;
458	ASMCpuId_ECX_EDX(iExt, &uECX2, &uEDX2);
459	CHECKVAL(uECX2, s.uECX, "%x");
460	CHECKVAL(uEDX2, s.uEDX, "%x");
461
462	uEAX2 = s.uEAX - 1;
463	uEBX2 = s.uEBX - 1;
464	uECX2 = s.uECX - 1;
465	uEDX2 = s.uEDX - 1;
466	ASMCpuId(iExt, &uEAX2, &uEBX2, &uECX2, &uEDX2);
467	CHECKVAL(uEAX2, s.uEAX, "%x");
468	CHECKVAL(uEBX2, s.uEBX, "%x");
469	CHECKVAL(uECX2, s.uECX, "%x");
470	CHECKVAL(uEDX2, s.uEDX, "%x");
471	}
472
473	/*
474	* Understandable output
475	*/
476	ASMCpuId(0x80000000, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
477	RTTestIPrintf(RTTESTLVL_ALWAYS,
478	"Ext Name: %.4s%.4s%.4s\n"
479	"Ext Supports: 0x80000000-%#010x\n",
480	&s.uEBX, &s.uEDX, &s.uECX, s.uEAX);
481
482	if (cExtFunctions >= 0x80000001)
483	{
484	ASMCpuId(0x80000001, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
485	RTTestIPrintf(RTTESTLVL_ALWAYS,
486	"Family: %#x \tExtended: %#x \tEffective: %#x\n"
487	"Model: %#x \tExtended: %#x \tEffective: %#x\n"
488	"Stepping: %d\n"
489	"Brand ID: %#05x\n",
490	(s.uEAX >> 8) & 0xf, (s.uEAX >> 20) & 0x7f, ASMGetCpuFamily(s.uEAX),
491	(s.uEAX >> 4) & 0xf, (s.uEAX >> 16) & 0x0f, ASMGetCpuModel(s.uEAX, fIntel),
492	ASMGetCpuStepping(s.uEAX),
493	s.uEBX & 0xfff);
494
495	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features EDX: ");
496	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FPU");
497	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VME");
498	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DE");
499	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE");
500	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TSC");
501	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MSR");
502	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAE");
503	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCE");
504	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMPXCHG8B");
505	if (s.uEDX & RT_BIT(9)) RTTestIPrintf(RTTESTLVL_ALWAYS, " APIC");
506	if (s.uEDX & RT_BIT(10)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 10");
507	if (s.uEDX & RT_BIT(11)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SysCallSysRet");
508	if (s.uEDX & RT_BIT(12)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MTRR");
509	if (s.uEDX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PGE");
510	if (s.uEDX & RT_BIT(14)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCA");
511	if (s.uEDX & RT_BIT(15)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMOV");
512	if (s.uEDX & RT_BIT(16)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAT");
513	if (s.uEDX & RT_BIT(17)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE36");
514	if (s.uEDX & RT_BIT(18)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 18");
515	if (s.uEDX & RT_BIT(19)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 19");
516	if (s.uEDX & RT_BIT(20)) RTTestIPrintf(RTTESTLVL_ALWAYS, " NX");
517	if (s.uEDX & RT_BIT(21)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 21");
518	if (s.uEDX & RT_BIT(22)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MmxExt");
519	if (s.uEDX & RT_BIT(23)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MMX");
520	if (s.uEDX & RT_BIT(24)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FXSR");
521	if (s.uEDX & RT_BIT(25)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FastFXSR");
522	if (s.uEDX & RT_BIT(26)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 26");
523	if (s.uEDX & RT_BIT(27)) RTTestIPrintf(RTTESTLVL_ALWAYS, " RDTSCP");
524	if (s.uEDX & RT_BIT(28)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 28");
525	if (s.uEDX & RT_BIT(29)) RTTestIPrintf(RTTESTLVL_ALWAYS, " LongMode");
526	if (s.uEDX & RT_BIT(30)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3DNowExt");
527	if (s.uEDX & RT_BIT(31)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3DNow");
528	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
529
530	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features ECX: ");
531	if (s.uECX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " LahfSahf");
532	if (s.uECX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CmpLegacy");
533	if (s.uECX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SVM");
534	if (s.uECX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3");
535	if (s.uECX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " AltMovCr8");
536	for (iBit = 5; iBit < 32; iBit++)
537	if (s.uECX & RT_BIT(iBit))
538	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
539	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
540	}
541
542	char szString[443+1] = {0};
543	if (cExtFunctions >= 0x80000002)
544	ASMCpuId(0x80000002, &szString[0 + 0], &szString[0 + 4], &szString[0 + 8], &szString[0 + 12]);
545	if (cExtFunctions >= 0x80000003)
546	ASMCpuId(0x80000003, &szString[16 + 0], &szString[16 + 4], &szString[16 + 8], &szString[16 + 12]);
547	if (cExtFunctions >= 0x80000004)
548	ASMCpuId(0x80000004, &szString[32 + 0], &szString[32 + 4], &szString[32 + 8], &szString[32 + 12]);
549	if (cExtFunctions >= 0x80000002)
550	RTTestIPrintf(RTTESTLVL_ALWAYS, "Full Name: %s\n", szString);
551
552	if (cExtFunctions >= 0x80000005)
553	{
554	ASMCpuId(0x80000005, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
555	RTTestIPrintf(RTTESTLVL_ALWAYS,
556	"TLB 2/4M Instr/Uni: %s %3d entries\n"
557	"TLB 2/4M Data: %s %3d entries\n",
558	getCacheAss((s.uEAX >> 8) & 0xff), (s.uEAX >> 0) & 0xff,
559	getCacheAss((s.uEAX >> 24) & 0xff), (s.uEAX >> 16) & 0xff);
560	RTTestIPrintf(RTTESTLVL_ALWAYS,
561	"TLB 4K Instr/Uni: %s %3d entries\n"
562	"TLB 4K Data: %s %3d entries\n",
563	getCacheAss((s.uEBX >> 8) & 0xff), (s.uEBX >> 0) & 0xff,
564	getCacheAss((s.uEBX >> 24) & 0xff), (s.uEBX >> 16) & 0xff);
565	RTTestIPrintf(RTTESTLVL_ALWAYS,
566	"L1 Instr Cache Line Size: %d bytes\n"
567	"L1 Instr Cache Lines Per Tag: %d\n"
568	"L1 Instr Cache Associativity: %s\n"
569	"L1 Instr Cache Size: %d KB\n",
570	(s.uEDX >> 0) & 0xff,
571	(s.uEDX >> 8) & 0xff,
572	getCacheAss((s.uEDX >> 16) & 0xff),
573	(s.uEDX >> 24) & 0xff);
574	RTTestIPrintf(RTTESTLVL_ALWAYS,
575	"L1 Data Cache Line Size: %d bytes\n"
576	"L1 Data Cache Lines Per Tag: %d\n"
577	"L1 Data Cache Associativity: %s\n"
578	"L1 Data Cache Size: %d KB\n",
579	(s.uECX >> 0) & 0xff,
580	(s.uECX >> 8) & 0xff,
581	getCacheAss((s.uECX >> 16) & 0xff),
582	(s.uECX >> 24) & 0xff);
583	}
584
585	if (cExtFunctions >= 0x80000006)
586	{
587	ASMCpuId(0x80000006, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
588	RTTestIPrintf(RTTESTLVL_ALWAYS,
589	"L2 TLB 2/4M Instr/Uni: %s %4d entries\n"
590	"L2 TLB 2/4M Data: %s %4d entries\n",
591	getL2CacheAss((s.uEAX >> 12) & 0xf), (s.uEAX >> 0) & 0xfff,
592	getL2CacheAss((s.uEAX >> 28) & 0xf), (s.uEAX >> 16) & 0xfff);
593	RTTestIPrintf(RTTESTLVL_ALWAYS,
594	"L2 TLB 4K Instr/Uni: %s %4d entries\n"
595	"L2 TLB 4K Data: %s %4d entries\n",
596	getL2CacheAss((s.uEBX >> 12) & 0xf), (s.uEBX >> 0) & 0xfff,
597	getL2CacheAss((s.uEBX >> 28) & 0xf), (s.uEBX >> 16) & 0xfff);
598	RTTestIPrintf(RTTESTLVL_ALWAYS,
599	"L2 Cache Line Size: %d bytes\n"
600	"L2 Cache Lines Per Tag: %d\n"
601	"L2 Cache Associativity: %s\n"
602	"L2 Cache Size: %d KB\n",
603	(s.uEDX >> 0) & 0xff,
604	(s.uEDX >> 8) & 0xf,
605	getL2CacheAss((s.uEDX >> 12) & 0xf),
606	(s.uEDX >> 16) & 0xffff);
607	}
608
609	if (cExtFunctions >= 0x80000007)
610	{
611	ASMCpuId(0x80000007, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
612	RTTestIPrintf(RTTESTLVL_ALWAYS, "APM Features: ");
613	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TS");
614	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FID");
615	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VID");
616	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TTP");
617	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TM");
618	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " STC");
619	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 6");
620	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 7");
621	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TscInvariant");
622	for (iBit = 9; iBit < 32; iBit++)
623	if (s.uEDX & RT_BIT(iBit))
624	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
625	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
626	}
627
628	if (cExtFunctions >= 0x80000008)
629	{
630	ASMCpuId(0x80000008, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
631	RTTestIPrintf(RTTESTLVL_ALWAYS,
632	"Physical Address Width: %d bits\n"
633	"Virtual Address Width: %d bits\n"
634	"Guest Physical Address Width: %d bits\n",
635	(s.uEAX >> 0) & 0xff,
636	(s.uEAX >> 8) & 0xff,
637	(s.uEAX >> 16) & 0xff);
638	RTTestIPrintf(RTTESTLVL_ALWAYS,
639	"Physical Core Count: %d\n",
640	((s.uECX >> 0) & 0xff) + 1);
641	if ((s.uECX >> 12) & 0xf)
642	RTTestIPrintf(RTTESTLVL_ALWAYS, "ApicIdCoreIdSize: %d bits\n", (s.uECX >> 12) & 0xf);
643	}
644
645	if (cExtFunctions >= 0x8000000a)
646	{
647	ASMCpuId(0x8000000a, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
648	RTTestIPrintf(RTTESTLVL_ALWAYS,
649	"SVM Revision: %d (%#x)\n"
650	"Number of Address Space IDs: %d (%#x)\n",
651	s.uEAX & 0xff, s.uEAX & 0xff,
652	s.uEBX, s.uEBX);
653	}
654	}
655
656	# if 0
657	static void bruteForceCpuId(void)
658	{
659	RTTestISub("brute force CPUID leafs");
660	uint32_t auPrevValues[4] = { 0, 0, 0, 0};
661	uint32_t uLeaf = 0;
662	do
663	{
664	uint32_t auValues[4];
665	ASMCpuIdExSlow(uLeaf, 0, 0, 0, &auValues[0], &auValues[1], &auValues[2], &auValues[3]);
666	if ( (auValues[0] != auPrevValues[0] && auValues[0] != uLeaf)
667	\|\| (auValues[1] != auPrevValues[1] && auValues[1] != 0)
668	\|\| (auValues[2] != auPrevValues[2] && auValues[2] != 0)
669	\|\| (auValues[3] != auPrevValues[3] && auValues[3] != 0)
670	\|\| (uLeaf & (UINT32_C(0x08000000) - UINT32_C(1))) == 0)
671	{
672	RTTestIPrintf(RTTESTLVL_ALWAYS,
673	"%08x: %08x %08x %08x %08x\n", uLeaf,
674	auValues[0], auValues[1], auValues[2], auValues[3]);
675	}
676	auPrevValues[0] = auValues[0];
677	auPrevValues[1] = auValues[1];
678	auPrevValues[2] = auValues[2];
679	auPrevValues[3] = auValues[3];
680
681	//uint32_t uSubLeaf = 0;
682	//do
683	//{
684	//
685	//
686	//} while (false);
687	} while (uLeaf++ < UINT32_MAX);
688	}
689	# endif
690
691	#endif /* AMD64 \|\| X86 */
692
693	DECLINLINE(void) tstASMAtomicXchgU8Worker(uint8_t volatile *pu8)
694	{
695	*pu8 = 0;
696	CHECKOP(ASMAtomicXchgU8(pu8, 1), 0, "%#x", uint8_t);
697	CHECKVAL(*pu8, 1, "%#x");
698
699	CHECKOP(ASMAtomicXchgU8(pu8, 0), 1, "%#x", uint8_t);
700	CHECKVAL(*pu8, 0, "%#x");
701
702	CHECKOP(ASMAtomicXchgU8(pu8, UINT8_C(0xff)), 0, "%#x", uint8_t);
703	CHECKVAL(*pu8, 0xff, "%#x");
704
705	CHECKOP(ASMAtomicXchgU8(pu8, UINT8_C(0x87)), UINT8_C(0xff), "%#x", uint8_t);
706	CHECKVAL(*pu8, 0x87, "%#x");
707	}
708
709
710	static void tstASMAtomicXchgU8(void)
711	{
712	DO_SIMPLE_TEST(ASMAtomicXchgU8, uint8_t);
713	}
714
715
716	DECLINLINE(void) tstASMAtomicXchgU16Worker(uint16_t volatile *pu16)
717	{
718	*pu16 = 0;
719
720	CHECKOP(ASMAtomicXchgU16(pu16, 1), 0, "%#x", uint16_t);
721	CHECKVAL(*pu16, 1, "%#x");
722
723	CHECKOP(ASMAtomicXchgU16(pu16, 0), 1, "%#x", uint16_t);
724	CHECKVAL(*pu16, 0, "%#x");
725
726	CHECKOP(ASMAtomicXchgU16(pu16, 0xffff), 0, "%#x", uint16_t);
727	CHECKVAL(*pu16, 0xffff, "%#x");
728
729	CHECKOP(ASMAtomicXchgU16(pu16, 0x8765), 0xffff, "%#x", uint16_t);
730	CHECKVAL(*pu16, 0x8765, "%#x");
731	}
732
733
734	static void tstASMAtomicXchgU16(void)
735	{
736	DO_SIMPLE_TEST(ASMAtomicXchgU16, uint16_t);
737	}
738
739
740	DECLINLINE(void) tstASMAtomicXchgU32Worker(uint32_t volatile *pu32)
741	{
742	*pu32 = 0;
743
744	CHECKOP(ASMAtomicXchgU32(pu32, 1), 0, "%#x", uint32_t);
745	CHECKVAL(*pu32, 1, "%#x");
746
747	CHECKOP(ASMAtomicXchgU32(pu32, 0), 1, "%#x", uint32_t);
748	CHECKVAL(*pu32, 0, "%#x");
749
750	CHECKOP(ASMAtomicXchgU32(pu32, ~UINT32_C(0)), 0, "%#x", uint32_t);
751	CHECKVAL(*pu32, ~UINT32_C(0), "%#x");
752
753	CHECKOP(ASMAtomicXchgU32(pu32, 0x87654321), ~UINT32_C(0), "%#x", uint32_t);
754	CHECKVAL(*pu32, 0x87654321, "%#x");
755	}
756
757
758	static void tstASMAtomicXchgU32(void)
759	{
760	DO_SIMPLE_TEST(ASMAtomicXchgU32, uint32_t);
761	}
762
763
764	DECLINLINE(void) tstASMAtomicXchgU64Worker(uint64_t volatile *pu64)
765	{
766	*pu64 = 0;
767
768	CHECKOP(ASMAtomicXchgU64(pu64, 1), UINT64_C(0), "%#llx", uint64_t);
769	CHECKVAL(*pu64, UINT64_C(1), "%#llx");
770
771	CHECKOP(ASMAtomicXchgU64(pu64, 0), UINT64_C(1), "%#llx", uint64_t);
772	CHECKVAL(*pu64, UINT64_C(0), "%#llx");
773
774	CHECKOP(ASMAtomicXchgU64(pu64, ~UINT64_C(0)), UINT64_C(0), "%#llx", uint64_t);
775	CHECKVAL(*pu64, ~UINT64_C(0), "%#llx");
776
777	CHECKOP(ASMAtomicXchgU64(pu64, UINT64_C(0xfedcba0987654321)), ~UINT64_C(0), "%#llx", uint64_t);
778	CHECKVAL(*pu64, UINT64_C(0xfedcba0987654321), "%#llx");
779	}
780
781
782	static void tstASMAtomicXchgU64(void)
783	{
784	DO_SIMPLE_TEST(ASMAtomicXchgU64, uint64_t);
785	}
786
787
788	DECLINLINE(void) tstASMAtomicXchgPtrWorker(void * volatile *ppv)
789	{
790	*ppv = NULL;
791
792	CHECKOP(ASMAtomicXchgPtr(ppv, (void )(~(uintptr_t)0)), NULL, "%p", void );
793	CHECKVAL(ppv, (void )(~(uintptr_t)0), "%p");
794
795	CHECKOP(ASMAtomicXchgPtr(ppv, (void )(uintptr_t)0x87654321), (void )(~(uintptr_t)0), "%p", void *);
796	CHECKVAL(ppv, (void )(uintptr_t)0x87654321, "%p");
797
798	CHECKOP(ASMAtomicXchgPtr(ppv, NULL), (void )(uintptr_t)0x87654321, "%p", void );
799	CHECKVAL(*ppv, NULL, "%p");
800	}
801
802
803	static void tstASMAtomicXchgPtr(void)
804	{
805	DO_SIMPLE_TEST(ASMAtomicXchgPtr, void *);
806	}
807
808
809	DECLINLINE(void) tstASMAtomicCmpXchgU8Worker(uint8_t volatile *pu8)
810	{
811	*pu8 = 0xff;
812
813	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0, 0), false, "%d", bool);
814	CHECKVAL(*pu8, 0xff, "%x");
815
816	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0, 0xff), true, "%d", bool);
817	CHECKVAL(*pu8, 0, "%x");
818
819	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0x79, 0xff), false, "%d", bool);
820	CHECKVAL(*pu8, 0, "%x");
821
822	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0x97, 0), true, "%d", bool);
823	CHECKVAL(*pu8, 0x97, "%x");
824	}
825
826
827	static void tstASMAtomicCmpXchgU8(void)
828	{
829	DO_SIMPLE_TEST(ASMAtomicCmpXchgU8, uint8_t);
830	}
831
832
833	DECLINLINE(void) tstASMAtomicCmpXchgU32Worker(uint32_t volatile *pu32)
834	{
835	*pu32 = UINT32_C(0xffffffff);
836
837	CHECKOP(ASMAtomicCmpXchgU32(pu32, 0, 0), false, "%d", bool);
838	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
839
840	CHECKOP(ASMAtomicCmpXchgU32(pu32, 0, UINT32_C(0xffffffff)), true, "%d", bool);
841	CHECKVAL(*pu32, 0, "%x");
842
843	CHECKOP(ASMAtomicCmpXchgU32(pu32, UINT32_C(0x8008efd), UINT32_C(0xffffffff)), false, "%d", bool);
844	CHECKVAL(*pu32, 0, "%x");
845
846	CHECKOP(ASMAtomicCmpXchgU32(pu32, UINT32_C(0x8008efd), 0), true, "%d", bool);
847	CHECKVAL(*pu32, UINT32_C(0x8008efd), "%x");
848	}
849
850
851	static void tstASMAtomicCmpXchgU32(void)
852	{
853	DO_SIMPLE_TEST(ASMAtomicCmpXchgU32, uint32_t);
854	}
855
856
857
858	DECLINLINE(void) tstASMAtomicCmpXchgU64Worker(uint64_t volatile *pu64)
859	{
860	*pu64 = UINT64_C(0xffffffffffffff);
861
862	CHECKOP(ASMAtomicCmpXchgU64(pu64, 0, 0), false, "%d", bool);
863	CHECKVAL(*pu64, UINT64_C(0xffffffffffffff), "%#llx");
864
865	CHECKOP(ASMAtomicCmpXchgU64(pu64, 0, UINT64_C(0xffffffffffffff)), true, "%d", bool);
866	CHECKVAL(*pu64, 0, "%x");
867
868	CHECKOP(ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), UINT64_C(0xffffffff)), false, "%d", bool);
869	CHECKVAL(*pu64, 0, "%x");
870
871	CHECKOP(ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), UINT64_C(0xffffffff00000000)), false, "%d", bool);
872	CHECKVAL(*pu64, 0, "%x");
873
874	CHECKOP(ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), 0), true, "%d", bool);
875	CHECKVAL(*pu64, UINT64_C(0x80040008008efd), "%#llx");
876	}
877
878
879	static void tstASMAtomicCmpXchgU64(void)
880	{
881	DO_SIMPLE_TEST(ASMAtomicCmpXchgU64, uint64_t);
882	}
883
884
885	DECLINLINE(void) tstASMAtomicCmpXchgExU32Worker(uint32_t volatile *pu32)
886	{
887	*pu32 = UINT32_C(0xffffffff);
888	uint32_t u32Old = UINT32_C(0x80005111);
889
890	CHECKOP(ASMAtomicCmpXchgExU32(pu32, 0, 0, &u32Old), false, "%d", bool);
891	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
892	CHECKVAL(u32Old, UINT32_C(0xffffffff), "%x");
893
894	CHECKOP(ASMAtomicCmpXchgExU32(pu32, 0, UINT32_C(0xffffffff), &u32Old), true, "%d", bool);
895	CHECKVAL(*pu32, 0, "%x");
896	CHECKVAL(u32Old, UINT32_C(0xffffffff), "%x");
897
898	CHECKOP(ASMAtomicCmpXchgExU32(pu32, UINT32_C(0x8008efd), UINT32_C(0xffffffff), &u32Old), false, "%d", bool);
899	CHECKVAL(*pu32, 0, "%x");
900	CHECKVAL(u32Old, 0, "%x");
901
902	CHECKOP(ASMAtomicCmpXchgExU32(pu32, UINT32_C(0x8008efd), 0, &u32Old), true, "%d", bool);
903	CHECKVAL(*pu32, UINT32_C(0x8008efd), "%x");
904	CHECKVAL(u32Old, 0, "%x");
905
906	CHECKOP(ASMAtomicCmpXchgExU32(pu32, 0, UINT32_C(0x8008efd), &u32Old), true, "%d", bool);
907	CHECKVAL(*pu32, 0, "%x");
908	CHECKVAL(u32Old, UINT32_C(0x8008efd), "%x");
909	}
910
911
912	static void tstASMAtomicCmpXchgExU32(void)
913	{
914	DO_SIMPLE_TEST(ASMAtomicCmpXchgExU32, uint32_t);
915	}
916
917
918	DECLINLINE(void) tstASMAtomicCmpXchgExU64Worker(uint64_t volatile *pu64)
919	{
920	*pu64 = UINT64_C(0xffffffffffffffff);
921	uint64_t u64Old = UINT64_C(0x8000000051111111);
922
923	CHECKOP(ASMAtomicCmpXchgExU64(pu64, 0, 0, &u64Old), false, "%d", bool);
924	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%llx");
925	CHECKVAL(u64Old, UINT64_C(0xffffffffffffffff), "%llx");
926
927	CHECKOP(ASMAtomicCmpXchgExU64(pu64, 0, UINT64_C(0xffffffffffffffff), &u64Old), true, "%d", bool);
928	CHECKVAL(*pu64, UINT64_C(0), "%llx");
929	CHECKVAL(u64Old, UINT64_C(0xffffffffffffffff), "%llx");
930
931	CHECKOP(ASMAtomicCmpXchgExU64(pu64, UINT64_C(0x80040008008efd), 0xffffffff, &u64Old), false, "%d", bool);
932	CHECKVAL(*pu64, UINT64_C(0), "%llx");
933	CHECKVAL(u64Old, UINT64_C(0), "%llx");
934
935	CHECKOP(ASMAtomicCmpXchgExU64(pu64, UINT64_C(0x80040008008efd), UINT64_C(0xffffffff00000000), &u64Old), false, "%d", bool);
936	CHECKVAL(*pu64, UINT64_C(0), "%llx");
937	CHECKVAL(u64Old, UINT64_C(0), "%llx");
938
939	CHECKOP(ASMAtomicCmpXchgExU64(pu64, UINT64_C(0x80040008008efd), 0, &u64Old), true, "%d", bool);
940	CHECKVAL(*pu64, UINT64_C(0x80040008008efd), "%llx");
941	CHECKVAL(u64Old, UINT64_C(0), "%llx");
942
943	CHECKOP(ASMAtomicCmpXchgExU64(pu64, 0, UINT64_C(0x80040008008efd), &u64Old), true, "%d", bool);
944	CHECKVAL(*pu64, UINT64_C(0), "%llx");
945	CHECKVAL(u64Old, UINT64_C(0x80040008008efd), "%llx");
946	}
947
948
949	static void tstASMAtomicCmpXchgExU64(void)
950	{
951	DO_SIMPLE_TEST(ASMAtomicCmpXchgExU64, uint64_t);
952	}
953
954
955	DECLINLINE(void) tstASMAtomicReadU64Worker(uint64_t volatile *pu64)
956	{
957	*pu64 = 0;
958
959	CHECKOP(ASMAtomicReadU64(pu64), UINT64_C(0), "%#llx", uint64_t);
960	CHECKVAL(*pu64, UINT64_C(0), "%#llx");
961
962	*pu64 = ~UINT64_C(0);
963	CHECKOP(ASMAtomicReadU64(pu64), ~UINT64_C(0), "%#llx", uint64_t);
964	CHECKVAL(*pu64, ~UINT64_C(0), "%#llx");
965
966	*pu64 = UINT64_C(0xfedcba0987654321);
967	CHECKOP(ASMAtomicReadU64(pu64), UINT64_C(0xfedcba0987654321), "%#llx", uint64_t);
968	CHECKVAL(*pu64, UINT64_C(0xfedcba0987654321), "%#llx");
969	}
970
971
972	static void tstASMAtomicReadU64(void)
973	{
974	DO_SIMPLE_TEST(ASMAtomicReadU64, uint64_t);
975	}
976
977
978	DECLINLINE(void) tstASMAtomicUoReadU64Worker(uint64_t volatile *pu64)
979	{
980	*pu64 = 0;
981
982	CHECKOP(ASMAtomicUoReadU64(pu64), UINT64_C(0), "%#llx", uint64_t);
983	CHECKVAL(*pu64, UINT64_C(0), "%#llx");
984
985	*pu64 = ~UINT64_C(0);
986	CHECKOP(ASMAtomicUoReadU64(pu64), ~UINT64_C(0), "%#llx", uint64_t);
987	CHECKVAL(*pu64, ~UINT64_C(0), "%#llx");
988
989	*pu64 = UINT64_C(0xfedcba0987654321);
990	CHECKOP(ASMAtomicUoReadU64(pu64), UINT64_C(0xfedcba0987654321), "%#llx", uint64_t);
991	CHECKVAL(*pu64, UINT64_C(0xfedcba0987654321), "%#llx");
992	}
993
994
995	static void tstASMAtomicUoReadU64(void)
996	{
997	DO_SIMPLE_TEST(ASMAtomicUoReadU64, uint64_t);
998	}
999
1000
1001	DECLINLINE(void) tstASMAtomicAddS32Worker(int32_t *pi32)
1002	{
1003	int32_t i32Rc;
1004	*pi32 = 10;
1005	#define MYCHECK(op, rc, val) \
1006	do { \
1007	i32Rc = op; \
1008	if (i32Rc != (rc)) \
1009	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %d expected %d\n", __FUNCTION__, __LINE__, #op, i32Rc, rc); \
1010	if (*pi32 != (val)) \
1011	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi32=%d expected %d\n", __FUNCTION__, __LINE__, #op, pi32, val); \
1012	} while (0)
1013	MYCHECK(ASMAtomicAddS32(pi32, 1), 10, 11);
1014	MYCHECK(ASMAtomicAddS32(pi32, -2), 11, 9);
1015	MYCHECK(ASMAtomicAddS32(pi32, -9), 9, 0);
1016	MYCHECK(ASMAtomicAddS32(pi32, -0x7fffffff), 0, -0x7fffffff);
1017	MYCHECK(ASMAtomicAddS32(pi32, 0), -0x7fffffff, -0x7fffffff);
1018	MYCHECK(ASMAtomicAddS32(pi32, 0x7fffffff), -0x7fffffff, 0);
1019	MYCHECK(ASMAtomicAddS32(pi32, 0), 0, 0);
1020	#undef MYCHECK
1021	}
1022
1023
1024	static void tstASMAtomicAddS32(void)
1025	{
1026	DO_SIMPLE_TEST(ASMAtomicAddS32, int32_t);
1027	}
1028
1029
1030	DECLINLINE(void) tstASMAtomicUoIncU32Worker(uint32_t volatile *pu32)
1031	{
1032	*pu32 = 0;
1033
1034	CHECKOP(ASMAtomicUoIncU32(pu32), UINT32_C(1), "%#x", uint32_t);
1035	CHECKVAL(*pu32, UINT32_C(1), "%#x");
1036
1037	*pu32 = ~UINT32_C(0);
1038	CHECKOP(ASMAtomicUoIncU32(pu32), 0, "%#x", uint32_t);
1039	CHECKVAL(*pu32, 0, "%#x");
1040
1041	*pu32 = UINT32_C(0x7fffffff);
1042	CHECKOP(ASMAtomicUoIncU32(pu32), UINT32_C(0x80000000), "%#x", uint32_t);
1043	CHECKVAL(*pu32, UINT32_C(0x80000000), "%#x");
1044	}
1045
1046
1047	static void tstASMAtomicUoIncU32(void)
1048	{
1049	DO_SIMPLE_TEST(ASMAtomicUoIncU32, uint32_t);
1050	}
1051
1052
1053	DECLINLINE(void) tstASMAtomicUoDecU32Worker(uint32_t volatile *pu32)
1054	{
1055	*pu32 = 0;
1056
1057	CHECKOP(ASMAtomicUoDecU32(pu32), ~UINT32_C(0), "%#x", uint32_t);
1058	CHECKVAL(*pu32, ~UINT32_C(0), "%#x");
1059
1060	*pu32 = ~UINT32_C(0);
1061	CHECKOP(ASMAtomicUoDecU32(pu32), UINT32_C(0xfffffffe), "%#x", uint32_t);
1062	CHECKVAL(*pu32, UINT32_C(0xfffffffe), "%#x");
1063
1064	*pu32 = UINT32_C(0x80000000);
1065	CHECKOP(ASMAtomicUoDecU32(pu32), UINT32_C(0x7fffffff), "%#x", uint32_t);
1066	CHECKVAL(*pu32, UINT32_C(0x7fffffff), "%#x");
1067	}
1068
1069
1070	static void tstASMAtomicUoDecU32(void)
1071	{
1072	DO_SIMPLE_TEST(ASMAtomicUoDecU32, uint32_t);
1073	}
1074
1075
1076	DECLINLINE(void) tstASMAtomicAddS64Worker(int64_t volatile *pi64)
1077	{
1078	int64_t i64Rc;
1079	*pi64 = 10;
1080	#define MYCHECK(op, rc, val) \
1081	do { \
1082	i64Rc = op; \
1083	if (i64Rc != (rc)) \
1084	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %llx expected %llx\n", __FUNCTION__, __LINE__, #op, i64Rc, (int64_t)rc); \
1085	if (*pi64 != (val)) \
1086	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi64=%llx expected %llx\n", __FUNCTION__, __LINE__, #op, pi64, (int64_t)(val)); \
1087	} while (0)
1088	MYCHECK(ASMAtomicAddS64(pi64, 1), 10, 11);
1089	MYCHECK(ASMAtomicAddS64(pi64, -2), 11, 9);
1090	MYCHECK(ASMAtomicAddS64(pi64, -9), 9, 0);
1091	MYCHECK(ASMAtomicAddS64(pi64, -INT64_MAX), 0, -INT64_MAX);
1092	MYCHECK(ASMAtomicAddS64(pi64, 0), -INT64_MAX, -INT64_MAX);
1093	MYCHECK(ASMAtomicAddS64(pi64, -1), -INT64_MAX, INT64_MIN);
1094	MYCHECK(ASMAtomicAddS64(pi64, INT64_MAX), INT64_MIN, -1);
1095	MYCHECK(ASMAtomicAddS64(pi64, 1), -1, 0);
1096	MYCHECK(ASMAtomicAddS64(pi64, 0), 0, 0);
1097	#undef MYCHECK
1098	}
1099
1100
1101	static void tstASMAtomicAddS64(void)
1102	{
1103	DO_SIMPLE_TEST(ASMAtomicAddS64, int64_t);
1104	}
1105
1106
1107	DECLINLINE(void) tstASMAtomicDecIncS32Worker(int32_t volatile *pi32)
1108	{
1109	int32_t i32Rc;
1110	*pi32 = 10;
1111	#define MYCHECK(op, rc) \
1112	do { \
1113	i32Rc = op; \
1114	if (i32Rc != (rc)) \
1115	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %d expected %d\n", __FUNCTION__, __LINE__, #op, i32Rc, rc); \
1116	if (*pi32 != (rc)) \
1117	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi32=%d expected %d\n", __FUNCTION__, __LINE__, #op, pi32, rc); \
1118	} while (0)
1119	MYCHECK(ASMAtomicDecS32(pi32), 9);
1120	MYCHECK(ASMAtomicDecS32(pi32), 8);
1121	MYCHECK(ASMAtomicDecS32(pi32), 7);
1122	MYCHECK(ASMAtomicDecS32(pi32), 6);
1123	MYCHECK(ASMAtomicDecS32(pi32), 5);
1124	MYCHECK(ASMAtomicDecS32(pi32), 4);
1125	MYCHECK(ASMAtomicDecS32(pi32), 3);
1126	MYCHECK(ASMAtomicDecS32(pi32), 2);
1127	MYCHECK(ASMAtomicDecS32(pi32), 1);
1128	MYCHECK(ASMAtomicDecS32(pi32), 0);
1129	MYCHECK(ASMAtomicDecS32(pi32), -1);
1130	MYCHECK(ASMAtomicDecS32(pi32), -2);
1131	MYCHECK(ASMAtomicIncS32(pi32), -1);
1132	MYCHECK(ASMAtomicIncS32(pi32), 0);
1133	MYCHECK(ASMAtomicIncS32(pi32), 1);
1134	MYCHECK(ASMAtomicIncS32(pi32), 2);
1135	MYCHECK(ASMAtomicIncS32(pi32), 3);
1136	MYCHECK(ASMAtomicDecS32(pi32), 2);
1137	MYCHECK(ASMAtomicIncS32(pi32), 3);
1138	MYCHECK(ASMAtomicDecS32(pi32), 2);
1139	MYCHECK(ASMAtomicIncS32(pi32), 3);
1140	#undef MYCHECK
1141	}
1142
1143
1144	static void tstASMAtomicDecIncS32(void)
1145	{
1146	DO_SIMPLE_TEST(ASMAtomicDecIncS32, int32_t);
1147	}
1148
1149
1150	DECLINLINE(void) tstASMAtomicDecIncS64Worker(int64_t volatile *pi64)
1151	{
1152	int64_t i64Rc;
1153	*pi64 = 10;
1154	#define MYCHECK(op, rc) \
1155	do { \
1156	i64Rc = op; \
1157	if (i64Rc != (rc)) \
1158	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %lld expected %lld\n", __FUNCTION__, __LINE__, #op, i64Rc, rc); \
1159	if (*pi64 != (rc)) \
1160	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi64=%lld expected %lld\n", __FUNCTION__, __LINE__, #op, pi64, rc); \
1161	} while (0)
1162	MYCHECK(ASMAtomicDecS64(pi64), 9);
1163	MYCHECK(ASMAtomicDecS64(pi64), 8);
1164	MYCHECK(ASMAtomicDecS64(pi64), 7);
1165	MYCHECK(ASMAtomicDecS64(pi64), 6);
1166	MYCHECK(ASMAtomicDecS64(pi64), 5);
1167	MYCHECK(ASMAtomicDecS64(pi64), 4);
1168	MYCHECK(ASMAtomicDecS64(pi64), 3);
1169	MYCHECK(ASMAtomicDecS64(pi64), 2);
1170	MYCHECK(ASMAtomicDecS64(pi64), 1);
1171	MYCHECK(ASMAtomicDecS64(pi64), 0);
1172	MYCHECK(ASMAtomicDecS64(pi64), -1);
1173	MYCHECK(ASMAtomicDecS64(pi64), -2);
1174	MYCHECK(ASMAtomicIncS64(pi64), -1);
1175	MYCHECK(ASMAtomicIncS64(pi64), 0);
1176	MYCHECK(ASMAtomicIncS64(pi64), 1);
1177	MYCHECK(ASMAtomicIncS64(pi64), 2);
1178	MYCHECK(ASMAtomicIncS64(pi64), 3);
1179	MYCHECK(ASMAtomicDecS64(pi64), 2);
1180	MYCHECK(ASMAtomicIncS64(pi64), 3);
1181	MYCHECK(ASMAtomicDecS64(pi64), 2);
1182	MYCHECK(ASMAtomicIncS64(pi64), 3);
1183	#undef MYCHECK
1184	}
1185
1186
1187	static void tstASMAtomicDecIncS64(void)
1188	{
1189	DO_SIMPLE_TEST(ASMAtomicDecIncS64, int64_t);
1190	}
1191
1192
1193	DECLINLINE(void) tstASMAtomicAndOrU32Worker(uint32_t volatile *pu32)
1194	{
1195	*pu32 = UINT32_C(0xffffffff);
1196
1197	ASMAtomicOrU32(pu32, UINT32_C(0xffffffff));
1198	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
1199
1200	ASMAtomicAndU32(pu32, UINT32_C(0xffffffff));
1201	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
1202
1203	ASMAtomicAndU32(pu32, UINT32_C(0x8f8f8f8f));
1204	CHECKVAL(*pu32, UINT32_C(0x8f8f8f8f), "%x");
1205
1206	ASMAtomicOrU32(pu32, UINT32_C(0x70707070));
1207	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
1208
1209	ASMAtomicAndU32(pu32, UINT32_C(1));
1210	CHECKVAL(*pu32, UINT32_C(1), "%x");
1211
1212	ASMAtomicOrU32(pu32, UINT32_C(0x80000000));
1213	CHECKVAL(*pu32, UINT32_C(0x80000001), "%x");
1214
1215	ASMAtomicAndU32(pu32, UINT32_C(0x80000000));
1216	CHECKVAL(*pu32, UINT32_C(0x80000000), "%x");
1217
1218	ASMAtomicAndU32(pu32, UINT32_C(0));
1219	CHECKVAL(*pu32, UINT32_C(0), "%x");
1220
1221	ASMAtomicOrU32(pu32, UINT32_C(0x42424242));
1222	CHECKVAL(*pu32, UINT32_C(0x42424242), "%x");
1223	}
1224
1225
1226	static void tstASMAtomicAndOrU32(void)
1227	{
1228	DO_SIMPLE_TEST(ASMAtomicAndOrU32, uint32_t);
1229	}
1230
1231
1232	DECLINLINE(void) tstASMAtomicAndOrU64Worker(uint64_t volatile *pu64)
1233	{
1234	*pu64 = UINT64_C(0xffffffff);
1235
1236	ASMAtomicOrU64(pu64, UINT64_C(0xffffffff));
1237	CHECKVAL(*pu64, UINT64_C(0xffffffff), "%x");
1238
1239	ASMAtomicAndU64(pu64, UINT64_C(0xffffffff));
1240	CHECKVAL(*pu64, UINT64_C(0xffffffff), "%x");
1241
1242	ASMAtomicAndU64(pu64, UINT64_C(0x8f8f8f8f));
1243	CHECKVAL(*pu64, UINT64_C(0x8f8f8f8f), "%x");
1244
1245	ASMAtomicOrU64(pu64, UINT64_C(0x70707070));
1246	CHECKVAL(*pu64, UINT64_C(0xffffffff), "%x");
1247
1248	ASMAtomicAndU64(pu64, UINT64_C(1));
1249	CHECKVAL(*pu64, UINT64_C(1), "%x");
1250
1251	ASMAtomicOrU64(pu64, UINT64_C(0x80000000));
1252	CHECKVAL(*pu64, UINT64_C(0x80000001), "%x");
1253
1254	ASMAtomicAndU64(pu64, UINT64_C(0x80000000));
1255	CHECKVAL(*pu64, UINT64_C(0x80000000), "%x");
1256
1257	ASMAtomicAndU64(pu64, UINT64_C(0));
1258	CHECKVAL(*pu64, UINT64_C(0), "%x");
1259
1260	ASMAtomicOrU64(pu64, UINT64_C(0x42424242));
1261	CHECKVAL(*pu64, UINT64_C(0x42424242), "%x");
1262
1263	// Same as above, but now 64-bit wide.
1264	ASMAtomicAndU64(pu64, UINT64_C(0));
1265	CHECKVAL(*pu64, UINT64_C(0), "%x");
1266
1267	ASMAtomicOrU64(pu64, UINT64_C(0xffffffffffffffff));
1268	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%x");
1269
1270	ASMAtomicAndU64(pu64, UINT64_C(0xffffffffffffffff));
1271	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%x");
1272
1273	ASMAtomicAndU64(pu64, UINT64_C(0x8f8f8f8f8f8f8f8f));
1274	CHECKVAL(*pu64, UINT64_C(0x8f8f8f8f8f8f8f8f), "%x");
1275
1276	ASMAtomicOrU64(pu64, UINT64_C(0x7070707070707070));
1277	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%x");
1278
1279	ASMAtomicAndU64(pu64, UINT64_C(1));
1280	CHECKVAL(*pu64, UINT64_C(1), "%x");
1281
1282	ASMAtomicOrU64(pu64, UINT64_C(0x8000000000000000));
1283	CHECKVAL(*pu64, UINT64_C(0x8000000000000001), "%x");
1284
1285	ASMAtomicAndU64(pu64, UINT64_C(0x8000000000000000));
1286	CHECKVAL(*pu64, UINT64_C(0x8000000000000000), "%x");
1287
1288	ASMAtomicAndU64(pu64, UINT64_C(0));
1289	CHECKVAL(*pu64, UINT64_C(0), "%x");
1290
1291	ASMAtomicOrU64(pu64, UINT64_C(0x4242424242424242));
1292	CHECKVAL(*pu64, UINT64_C(0x4242424242424242), "%x");
1293	}
1294
1295
1296	static void tstASMAtomicAndOrU64(void)
1297	{
1298	DO_SIMPLE_TEST(ASMAtomicAndOrU64, uint64_t);
1299	}
1300
1301
1302	DECLINLINE(void) tstASMAtomicUoAndOrU32Worker(uint32_t volatile *pu32)
1303	{
1304	*pu32 = UINT32_C(0xffffffff);
1305
1306	ASMAtomicUoOrU32(pu32, UINT32_C(0xffffffff));
1307	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%#x");
1308
1309	ASMAtomicUoAndU32(pu32, UINT32_C(0xffffffff));
1310	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%#x");
1311
1312	ASMAtomicUoAndU32(pu32, UINT32_C(0x8f8f8f8f));
1313	CHECKVAL(*pu32, UINT32_C(0x8f8f8f8f), "%#x");
1314
1315	ASMAtomicUoOrU32(pu32, UINT32_C(0x70707070));
1316	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%#x");
1317
1318	ASMAtomicUoAndU32(pu32, UINT32_C(1));
1319	CHECKVAL(*pu32, UINT32_C(1), "%#x");
1320
1321	ASMAtomicUoOrU32(pu32, UINT32_C(0x80000000));
1322	CHECKVAL(*pu32, UINT32_C(0x80000001), "%#x");
1323
1324	ASMAtomicUoAndU32(pu32, UINT32_C(0x80000000));
1325	CHECKVAL(*pu32, UINT32_C(0x80000000), "%#x");
1326
1327	ASMAtomicUoAndU32(pu32, UINT32_C(0));
1328	CHECKVAL(*pu32, UINT32_C(0), "%#x");
1329
1330	ASMAtomicUoOrU32(pu32, UINT32_C(0x42424242));
1331	CHECKVAL(*pu32, UINT32_C(0x42424242), "%#x");
1332	}
1333
1334
1335	static void tstASMAtomicUoAndOrU32(void)
1336	{
1337	DO_SIMPLE_TEST(ASMAtomicUoAndOrU32, uint32_t);
1338	}
1339
1340
1341	typedef struct
1342	{
1343	uint8_t ab[PAGE_SIZE];
1344	} TSTPAGE;
1345
1346
1347	DECLINLINE(void) tstASMMemZeroPageWorker(TSTPAGE *pPage)
1348	{
1349	for (unsigned j = 0; j < 16; j++)
1350	{
1351	memset(pPage, 0x11 * j, sizeof(*pPage));
1352	ASMMemZeroPage(pPage);
1353	for (unsigned i = 0; i < sizeof(pPage->ab); i++)
1354	if (pPage->ab[i])
1355	RTTestFailed(g_hTest, "ASMMemZeroPage didn't clear byte at offset %#x!\n", i);
1356	}
1357	}
1358
1359
1360	static void tstASMMemZeroPage(void)
1361	{
1362	DO_SIMPLE_TEST(ASMMemZeroPage, TSTPAGE);
1363	}
1364
1365
1366	void tstASMMemIsZeroPage(RTTEST hTest)
1367	{
1368	RTTestSub(hTest, "ASMMemIsZeroPage");
1369
1370	void *pvPage1 = RTTestGuardedAllocHead(hTest, PAGE_SIZE);
1371	void *pvPage2 = RTTestGuardedAllocTail(hTest, PAGE_SIZE);
1372	RTTESTI_CHECK_RETV(pvPage1 && pvPage2);
1373
1374	memset(pvPage1, 0, PAGE_SIZE);
1375	memset(pvPage2, 0, PAGE_SIZE);
1376	RTTESTI_CHECK(ASMMemIsZeroPage(pvPage1));
1377	RTTESTI_CHECK(ASMMemIsZeroPage(pvPage2));
1378
1379	memset(pvPage1, 0xff, PAGE_SIZE);
1380	memset(pvPage2, 0xff, PAGE_SIZE);
1381	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage1));
1382	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage2));
1383
1384	memset(pvPage1, 0, PAGE_SIZE);
1385	memset(pvPage2, 0, PAGE_SIZE);
1386	for (unsigned off = 0; off < PAGE_SIZE; off++)
1387	{
1388	((uint8_t *)pvPage1)[off] = 1;
1389	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage1));
1390	((uint8_t *)pvPage1)[off] = 0;
1391
1392	((uint8_t *)pvPage2)[off] = 0x80;
1393	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage2));
1394	((uint8_t *)pvPage2)[off] = 0;
1395	}
1396
1397	RTTestSubDone(hTest);
1398	}
1399
1400
1401	void tstASMMemFirstMismatchingU8(RTTEST hTest)
1402	{
1403	RTTestSub(hTest, "ASMMemFirstMismatchingU8");
1404
1405	uint8_t pbPage1 = (uint8_t )RTTestGuardedAllocHead(hTest, PAGE_SIZE);
1406	uint8_t pbPage2 = (uint8_t )RTTestGuardedAllocTail(hTest, PAGE_SIZE);
1407	RTTESTI_CHECK_RETV(pbPage1 && pbPage2);
1408
1409	memset(pbPage1, 0, PAGE_SIZE);
1410	memset(pbPage2, 0, PAGE_SIZE);
1411	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0) == NULL);
1412	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0) == NULL);
1413	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 1) == pbPage1);
1414	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 1) == pbPage2);
1415	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0x87) == pbPage1);
1416	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0x87) == pbPage2);
1417	RTTESTI_CHECK(ASMMemIsZero(pbPage1, PAGE_SIZE));
1418	RTTESTI_CHECK(ASMMemIsZero(pbPage2, PAGE_SIZE));
1419	RTTESTI_CHECK(ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0));
1420	RTTESTI_CHECK(ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0));
1421	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0x34));
1422	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0x88));
1423	unsigned cbSub = 32;
1424	while (cbSub-- > 0)
1425	{
1426	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0) == NULL);
1427	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0) == NULL);
1428	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0) == NULL);
1429	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0) == NULL);
1430
1431	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0x34) == &pbPage1[PAGE_SIZE - cbSub] \|\| !cbSub);
1432	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0x99) == &pbPage2[PAGE_SIZE - cbSub] \|\| !cbSub);
1433	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0x42) == pbPage1 \|\| !cbSub);
1434	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0x88) == pbPage2 \|\| !cbSub);
1435	}
1436
1437	memset(pbPage1, 0xff, PAGE_SIZE);
1438	memset(pbPage2, 0xff, PAGE_SIZE);
1439	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0xff) == NULL);
1440	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0xff) == NULL);
1441	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0xfe) == pbPage1);
1442	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0xfe) == pbPage2);
1443	RTTESTI_CHECK(!ASMMemIsZero(pbPage1, PAGE_SIZE));
1444	RTTESTI_CHECK(!ASMMemIsZero(pbPage2, PAGE_SIZE));
1445	RTTESTI_CHECK(ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0xff));
1446	RTTESTI_CHECK(ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0xff));
1447	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0));
1448	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0));
1449	cbSub = 32;
1450	while (cbSub-- > 0)
1451	{
1452	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0xff) == NULL);
1453	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0xff) == NULL);
1454	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0xff) == NULL);
1455	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0xff) == NULL);
1456
1457	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0xfe) == &pbPage1[PAGE_SIZE - cbSub] \|\| !cbSub);
1458	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0xfe) == &pbPage2[PAGE_SIZE - cbSub] \|\| !cbSub);
1459	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0xfe) == pbPage1 \|\| !cbSub);
1460	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0xfe) == pbPage2 \|\| !cbSub);
1461	}
1462
1463
1464	/*
1465	* Various alignments and sizes.
1466	*/
1467	uint8_t const bFiller1 = 0x00;
1468	uint8_t const bFiller2 = 0xf6;
1469	size_t const cbBuf = 128;
1470	uint8_t *pbBuf1 = pbPage1;
1471	uint8_t pbBuf2 = &pbPage2[PAGE_SIZE - cbBuf]; / Put it up against the tail guard */
1472	memset(pbPage1, ~bFiller1, PAGE_SIZE);
1473	memset(pbPage2, ~bFiller2, PAGE_SIZE);
1474	memset(pbBuf1, bFiller1, cbBuf);
1475	memset(pbBuf2, bFiller2, cbBuf);
1476	for (size_t offNonZero = 0; offNonZero < cbBuf; offNonZero++)
1477	{
1478	uint8_t bRand = (uint8_t)RTRandU32();
1479	pbBuf1[offNonZero] = bRand \| 1;
1480	pbBuf2[offNonZero] = (0x80 \| bRand) ^ 0xf6;
1481
1482	for (size_t offStart = 0; offStart < 32; offStart++)
1483	{
1484	size_t const cbMax = cbBuf - offStart;
1485	for (size_t cb = 0; cb < cbMax; cb++)
1486	{
1487	size_t const offEnd = offStart + cb;
1488	uint8_t bSaved1, bSaved2;
1489	if (offEnd < PAGE_SIZE)
1490	{
1491	bSaved1 = pbBuf1[offEnd];
1492	bSaved2 = pbBuf2[offEnd];
1493	pbBuf1[offEnd] = 0xff;
1494	pbBuf2[offEnd] = 0xff;
1495	}
1496	#ifdef _MSC_VER /* simple stupid compiler warnings */
1497	else
1498	bSaved1 = bSaved2 = 0;
1499	#endif
1500
1501	uint8_t pbRet = (uint8_t )ASMMemFirstMismatchingU8(pbBuf1 + offStart, cb, bFiller1);
1502	RTTESTI_CHECK(offNonZero - offStart < cb ? pbRet == &pbBuf1[offNonZero] : pbRet == NULL);
1503
1504	pbRet = (uint8_t *)ASMMemFirstMismatchingU8(pbBuf2 + offStart, cb, bFiller2);
1505	RTTESTI_CHECK(offNonZero - offStart < cb ? pbRet == &pbBuf2[offNonZero] : pbRet == NULL);
1506
1507	if (offEnd < PAGE_SIZE)
1508	{
1509	pbBuf1[offEnd] = bSaved1;
1510	pbBuf2[offEnd] = bSaved2;
1511	}
1512	}
1513	}
1514
1515	pbBuf1[offNonZero] = 0;
1516	pbBuf2[offNonZero] = 0xf6;
1517	}
1518
1519	RTTestSubDone(hTest);
1520	}
1521
1522
1523	void tstASMMemZero32(void)
1524	{
1525	RTTestSub(g_hTest, "ASMMemFill32");
1526
1527	struct
1528	{
1529	uint64_t u64Magic1;
1530	uint8_t abPage[PAGE_SIZE - 32];
1531	uint64_t u64Magic2;
1532	} Buf1, Buf2, Buf3;
1533
1534	Buf1.u64Magic1 = UINT64_C(0xffffffffffffffff);
1535	memset(Buf1.abPage, 0x55, sizeof(Buf1.abPage));
1536	Buf1.u64Magic2 = UINT64_C(0xffffffffffffffff);
1537	Buf2.u64Magic1 = UINT64_C(0xffffffffffffffff);
1538	memset(Buf2.abPage, 0x77, sizeof(Buf2.abPage));
1539	Buf2.u64Magic2 = UINT64_C(0xffffffffffffffff);
1540	Buf3.u64Magic1 = UINT64_C(0xffffffffffffffff);
1541	memset(Buf3.abPage, 0x99, sizeof(Buf3.abPage));
1542	Buf3.u64Magic2 = UINT64_C(0xffffffffffffffff);
1543	ASMMemZero32(Buf1.abPage, sizeof(Buf1.abPage));
1544	ASMMemZero32(Buf2.abPage, sizeof(Buf2.abPage));
1545	ASMMemZero32(Buf3.abPage, sizeof(Buf3.abPage));
1546	if ( Buf1.u64Magic1 != UINT64_C(0xffffffffffffffff)
1547	\|\| Buf1.u64Magic2 != UINT64_C(0xffffffffffffffff)
1548	\|\| Buf2.u64Magic1 != UINT64_C(0xffffffffffffffff)
1549	\|\| Buf2.u64Magic2 != UINT64_C(0xffffffffffffffff)
1550	\|\| Buf3.u64Magic1 != UINT64_C(0xffffffffffffffff)
1551	\|\| Buf3.u64Magic2 != UINT64_C(0xffffffffffffffff))
1552	{
1553	RTTestFailed(g_hTest, "ASMMemZero32 violated one/both magic(s)!\n");
1554	}
1555	for (unsigned i = 0; i < RT_ELEMENTS(Buf1.abPage); i++)
1556	if (Buf1.abPage[i])
1557	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
1558	for (unsigned i = 0; i < RT_ELEMENTS(Buf2.abPage); i++)
1559	if (Buf2.abPage[i])
1560	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
1561	for (unsigned i = 0; i < RT_ELEMENTS(Buf3.abPage); i++)
1562	if (Buf3.abPage[i])
1563	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
1564	}
1565
1566
1567	void tstASMMemFill32(void)
1568	{
1569	RTTestSub(g_hTest, "ASMMemFill32");
1570
1571	struct
1572	{
1573	uint64_t u64Magic1;
1574	uint32_t au32Page[PAGE_SIZE / 4];
1575	uint64_t u64Magic2;
1576	} Buf1;
1577	struct
1578	{
1579	uint64_t u64Magic1;
1580	uint32_t au32Page[(PAGE_SIZE / 4) - 3];
1581	uint64_t u64Magic2;
1582	} Buf2;
1583	struct
1584	{
1585	uint64_t u64Magic1;
1586	uint32_t au32Page[(PAGE_SIZE / 4) - 1];
1587	uint64_t u64Magic2;
1588	} Buf3;
1589
1590	Buf1.u64Magic1 = UINT64_C(0xffffffffffffffff);
1591	memset(Buf1.au32Page, 0x55, sizeof(Buf1.au32Page));
1592	Buf1.u64Magic2 = UINT64_C(0xffffffffffffffff);
1593	Buf2.u64Magic1 = UINT64_C(0xffffffffffffffff);
1594	memset(Buf2.au32Page, 0x77, sizeof(Buf2.au32Page));
1595	Buf2.u64Magic2 = UINT64_C(0xffffffffffffffff);
1596	Buf3.u64Magic1 = UINT64_C(0xffffffffffffffff);
1597	memset(Buf3.au32Page, 0x99, sizeof(Buf3.au32Page));
1598	Buf3.u64Magic2 = UINT64_C(0xffffffffffffffff);
1599	ASMMemFill32(Buf1.au32Page, sizeof(Buf1.au32Page), 0xdeadbeef);
1600	ASMMemFill32(Buf2.au32Page, sizeof(Buf2.au32Page), 0xcafeff01);
1601	ASMMemFill32(Buf3.au32Page, sizeof(Buf3.au32Page), 0xf00dd00f);
1602	if ( Buf1.u64Magic1 != UINT64_C(0xffffffffffffffff)
1603	\|\| Buf1.u64Magic2 != UINT64_C(0xffffffffffffffff)
1604	\|\| Buf2.u64Magic1 != UINT64_C(0xffffffffffffffff)
1605	\|\| Buf2.u64Magic2 != UINT64_C(0xffffffffffffffff)
1606	\|\| Buf3.u64Magic1 != UINT64_C(0xffffffffffffffff)
1607	\|\| Buf3.u64Magic2 != UINT64_C(0xffffffffffffffff))
1608	RTTestFailed(g_hTest, "ASMMemFill32 violated one/both magic(s)!\n");
1609	for (unsigned i = 0; i < RT_ELEMENTS(Buf1.au32Page); i++)
1610	if (Buf1.au32Page[i] != 0xdeadbeef)
1611	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf1.au32Page[i], 0xdeadbeef);
1612	for (unsigned i = 0; i < RT_ELEMENTS(Buf2.au32Page); i++)
1613	if (Buf2.au32Page[i] != 0xcafeff01)
1614	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf2.au32Page[i], 0xcafeff01);
1615	for (unsigned i = 0; i < RT_ELEMENTS(Buf3.au32Page); i++)
1616	if (Buf3.au32Page[i] != 0xf00dd00f)
1617	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf3.au32Page[i], 0xf00dd00f);
1618	}
1619
1620
1621
1622	void tstASMMath(void)
1623	{
1624	RTTestSub(g_hTest, "Math");
1625
1626	uint64_t u64 = ASMMult2xU32RetU64(UINT32_C(0x80000000), UINT32_C(0x10000000));
1627	CHECKVAL(u64, UINT64_C(0x0800000000000000), "%#018RX64");
1628
1629	uint32_t u32 = ASMDivU64ByU32RetU32(UINT64_C(0x0800000000000000), UINT32_C(0x10000000));
1630	CHECKVAL(u32, UINT32_C(0x80000000), "%#010RX32");
1631
1632	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x00000001), UINT32_C(0x00000001), UINT32_C(0x00000001));
1633	CHECKVAL(u32, UINT32_C(0x00000001), "%#018RX32");
1634	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x10000000), UINT32_C(0x80000000), UINT32_C(0x20000000));
1635	CHECKVAL(u32, UINT32_C(0x40000000), "%#018RX32");
1636	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x76543210), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
1637	CHECKVAL(u32, UINT32_C(0x76543210), "%#018RX32");
1638	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0xffffffff), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
1639	CHECKVAL(u32, UINT32_C(0xffffffff), "%#018RX32");
1640	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0xffffffff), UINT32_C(0xfffffff0), UINT32_C(0xffffffff));
1641	CHECKVAL(u32, UINT32_C(0xfffffff0), "%#018RX32");
1642	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x10359583), UINT32_C(0x58734981), UINT32_C(0xf8694045));
1643	CHECKVAL(u32, UINT32_C(0x05c584ce), "%#018RX32");
1644	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x10359583), UINT32_C(0xf8694045), UINT32_C(0x58734981));
1645	CHECKVAL(u32, UINT32_C(0x2d860795), "%#018RX32");
1646
1647	#if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1648	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x0000000000000001), UINT32_C(0x00000001), UINT32_C(0x00000001));
1649	CHECKVAL(u64, UINT64_C(0x0000000000000001), "%#018RX64");
1650	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x0000000100000000), UINT32_C(0x80000000), UINT32_C(0x00000002));
1651	CHECKVAL(u64, UINT64_C(0x4000000000000000), "%#018RX64");
1652	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xfedcba9876543210), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
1653	CHECKVAL(u64, UINT64_C(0xfedcba9876543210), "%#018RX64");
1654	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xffffffffffffffff), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
1655	CHECKVAL(u64, UINT64_C(0xffffffffffffffff), "%#018RX64");
1656	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xffffffffffffffff), UINT32_C(0xfffffff0), UINT32_C(0xffffffff));
1657	CHECKVAL(u64, UINT64_C(0xfffffff0fffffff0), "%#018RX64");
1658	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x3415934810359583), UINT32_C(0x58734981), UINT32_C(0xf8694045));
1659	CHECKVAL(u64, UINT64_C(0x128b9c3d43184763), "%#018RX64");
1660	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x3415934810359583), UINT32_C(0xf8694045), UINT32_C(0x58734981));
1661	CHECKVAL(u64, UINT64_C(0x924719355cd35a27), "%#018RX64");
1662
1663	# if 0 /* bird: question is whether this should trap or not:
1664	*
1665	* frank: Of course it must trap:
1666	*
1667	* 0xfffffff8 * 0x77d7daf8 = 0x77d7daf441412840
1668	*
1669	* During the following division, the quotient must fit into a 32-bit register.
1670	* Therefore the smallest valid divisor is
1671	*
1672	* (0x77d7daf441412840 >> 32) + 1 = 0x77d7daf5
1673	*
1674	* which is definitely greater than 0x3b9aca00.
1675	*
1676	* bird: No, the C version does not crash. So, the question is whether there's any
1677	* code depending on it not crashing.
1678	*
1679	* Of course the assembly versions of the code crash right now for the reasons you've
1680	* given, but the 32-bit MSC version does not crash.
1681	*
1682	* frank: The C version does not crash but delivers incorrect results for this case.
1683	* The reason is
1684	*
1685	* u.s.Hi = (unsigned long)(u64Hi / u32C);
1686	*
1687	* Here the division is actually 64-bit by 64-bit but the 64-bit result is truncated
1688	* to 32 bit. If using this (optimized and fast) function we should just be sure that
1689	* the operands are in a valid range.
1690	*/
1691	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xfffffff8c65d6731), UINT32_C(0x77d7daf8), UINT32_C(0x3b9aca00));
1692	CHECKVAL(u64, UINT64_C(0x02b8f9a2aa74e3dc), "%#018RX64");
1693	# endif
1694	#endif /* AMD64 \|\| X86 */
1695
1696	u32 = ASMModU64ByU32RetU32(UINT64_C(0x0ffffff8c65d6731), UINT32_C(0x77d7daf8));
1697	CHECKVAL(u32, UINT32_C(0x3B642451), "%#010RX32");
1698
1699	int32_t i32;
1700	i32 = ASMModS64ByS32RetS32(INT64_C(-11), INT32_C(-2));
1701	CHECKVAL(i32, INT32_C(-1), "%010RI32");
1702	i32 = ASMModS64ByS32RetS32(INT64_C(-11), INT32_C(2));
1703	CHECKVAL(i32, INT32_C(-1), "%010RI32");
1704	i32 = ASMModS64ByS32RetS32(INT64_C(11), INT32_C(-2));
1705	CHECKVAL(i32, INT32_C(1), "%010RI32");
1706
1707	i32 = ASMModS64ByS32RetS32(INT64_C(92233720368547758), INT32_C(2147483647));
1708	CHECKVAL(i32, INT32_C(2104533974), "%010RI32");
1709	i32 = ASMModS64ByS32RetS32(INT64_C(-92233720368547758), INT32_C(2147483647));
1710	CHECKVAL(i32, INT32_C(-2104533974), "%010RI32");
1711	}
1712
1713
1714	void tstASMByteSwap(void)
1715	{
1716	RTTestSub(g_hTest, "ASMByteSwap*");
1717
1718	uint64_t u64In = UINT64_C(0x0011223344556677);
1719	uint64_t u64Out = ASMByteSwapU64(u64In);
1720	CHECKVAL(u64In, UINT64_C(0x0011223344556677), "%#018RX64");
1721	CHECKVAL(u64Out, UINT64_C(0x7766554433221100), "%#018RX64");
1722	u64Out = ASMByteSwapU64(u64Out);
1723	CHECKVAL(u64Out, u64In, "%#018RX64");
1724	u64In = UINT64_C(0x0123456789abcdef);
1725	u64Out = ASMByteSwapU64(u64In);
1726	CHECKVAL(u64In, UINT64_C(0x0123456789abcdef), "%#018RX64");
1727	CHECKVAL(u64Out, UINT64_C(0xefcdab8967452301), "%#018RX64");
1728	u64Out = ASMByteSwapU64(u64Out);
1729	CHECKVAL(u64Out, u64In, "%#018RX64");
1730	u64In = 0;
1731	u64Out = ASMByteSwapU64(u64In);
1732	CHECKVAL(u64Out, u64In, "%#018RX64");
1733	u64In = UINT64_MAX;
1734	u64Out = ASMByteSwapU64(u64In);
1735	CHECKVAL(u64Out, u64In, "%#018RX64");
1736
1737	uint32_t u32In = UINT32_C(0x00112233);
1738	uint32_t u32Out = ASMByteSwapU32(u32In);
1739	CHECKVAL(u32In, UINT32_C(0x00112233), "%#010RX32");
1740	CHECKVAL(u32Out, UINT32_C(0x33221100), "%#010RX32");
1741	u32Out = ASMByteSwapU32(u32Out);
1742	CHECKVAL(u32Out, u32In, "%#010RX32");
1743	u32In = UINT32_C(0x12345678);
1744	u32Out = ASMByteSwapU32(u32In);
1745	CHECKVAL(u32In, UINT32_C(0x12345678), "%#010RX32");
1746	CHECKVAL(u32Out, UINT32_C(0x78563412), "%#010RX32");
1747	u32Out = ASMByteSwapU32(u32Out);
1748	CHECKVAL(u32Out, u32In, "%#010RX32");
1749	u32In = 0;
1750	u32Out = ASMByteSwapU32(u32In);
1751	CHECKVAL(u32Out, u32In, "%#010RX32");
1752	u32In = UINT32_MAX;
1753	u32Out = ASMByteSwapU32(u32In);
1754	CHECKVAL(u32Out, u32In, "%#010RX32");
1755
1756	uint16_t u16In = UINT16_C(0x0011);
1757	uint16_t u16Out = ASMByteSwapU16(u16In);
1758	CHECKVAL(u16In, UINT16_C(0x0011), "%#06RX16");
1759	CHECKVAL(u16Out, UINT16_C(0x1100), "%#06RX16");
1760	u16Out = ASMByteSwapU16(u16Out);
1761	CHECKVAL(u16Out, u16In, "%#06RX16");
1762	u16In = UINT16_C(0x1234);
1763	u16Out = ASMByteSwapU16(u16In);
1764	CHECKVAL(u16In, UINT16_C(0x1234), "%#06RX16");
1765	CHECKVAL(u16Out, UINT16_C(0x3412), "%#06RX16");
1766	u16Out = ASMByteSwapU16(u16Out);
1767	CHECKVAL(u16Out, u16In, "%#06RX16");
1768	u16In = 0;
1769	u16Out = ASMByteSwapU16(u16In);
1770	CHECKVAL(u16Out, u16In, "%#06RX16");
1771	u16In = UINT16_MAX;
1772	u16Out = ASMByteSwapU16(u16In);
1773	CHECKVAL(u16Out, u16In, "%#06RX16");
1774	}
1775
1776
1777	void tstASMBench(void)
1778	{
1779	/*
1780	* Make this static. We don't want to have this located on the stack.
1781	*/
1782	static uint8_t volatile s_u8;
1783	static int8_t volatile s_i8;
1784	static uint16_t volatile s_u16;
1785	static int16_t volatile s_i16;
1786	static uint32_t volatile s_u32;
1787	static int32_t volatile s_i32;
1788	static uint64_t volatile s_u64;
1789	static int64_t volatile s_i64;
1790	unsigned i;
1791	const unsigned cRounds = _2M; /* Must be multiple of 8 */
1792	uint64_t u64Elapsed;
1793
1794	RTTestSub(g_hTest, "Benchmarking");
1795
1796	#if 0 && !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
1797	# define BENCH(op, str) \
1798	do { \
1799	RTThreadYield(); \
1800	u64Elapsed = ASMReadTSC(); \
1801	for (i = cRounds; i > 0; i--) \
1802	op; \
1803	u64Elapsed = ASMReadTSC() - u64Elapsed; \
1804	RTTestValue(g_hTest, str, u64Elapsed / cRounds, RTTESTUNIT_TICKS_PER_CALL); \
1805	} while (0)
1806	#else
1807	# define BENCH(op, str) \
1808	do { \
1809	RTThreadYield(); \
1810	u64Elapsed = RTTimeNanoTS(); \
1811	for (i = cRounds / 8; i > 0; i--) \
1812	{ \
1813	op; \
1814	op; \
1815	op; \
1816	op; \
1817	op; \
1818	op; \
1819	op; \
1820	op; \
1821	} \
1822	u64Elapsed = RTTimeNanoTS() - u64Elapsed; \
1823	RTTestValue(g_hTest, str, u64Elapsed / cRounds, RTTESTUNIT_NS_PER_CALL); \
1824	} while (0)
1825	#endif
1826	#if (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)) && !defined(GCC44_32BIT_PIC)
1827	# define BENCH_TSC(op, str) \
1828	do { \
1829	RTThreadYield(); \
1830	u64Elapsed = ASMReadTSC(); \
1831	for (i = cRounds / 8; i > 0; i--) \
1832	{ \
1833	op; \
1834	op; \
1835	op; \
1836	op; \
1837	op; \
1838	op; \
1839	op; \
1840	op; \
1841	} \
1842	u64Elapsed = ASMReadTSC() - u64Elapsed; \
1843	RTTestValue(g_hTest, str, u64Elapsed / cRounds, /RTTESTUNIT_TICKS_PER_CALL/ RTTESTUNIT_NONE); \
1844	} while (0)
1845	#else
1846	# define BENCH_TSC(op, str) BENCH(op, str)
1847	#endif
1848
1849	BENCH(s_u32 = 0, "s_u32 = 0");
1850	BENCH(ASMAtomicUoReadU8(&s_u8), "ASMAtomicUoReadU8");
1851	BENCH(ASMAtomicUoReadS8(&s_i8), "ASMAtomicUoReadS8");
1852	BENCH(ASMAtomicUoReadU16(&s_u16), "ASMAtomicUoReadU16");
1853	BENCH(ASMAtomicUoReadS16(&s_i16), "ASMAtomicUoReadS16");
1854	BENCH(ASMAtomicUoReadU32(&s_u32), "ASMAtomicUoReadU32");
1855	BENCH(ASMAtomicUoReadS32(&s_i32), "ASMAtomicUoReadS32");
1856	BENCH(ASMAtomicUoReadU64(&s_u64), "ASMAtomicUoReadU64");
1857	BENCH(ASMAtomicUoReadS64(&s_i64), "ASMAtomicUoReadS64");
1858	BENCH(ASMAtomicReadU8(&s_u8), "ASMAtomicReadU8");
1859	BENCH(ASMAtomicReadS8(&s_i8), "ASMAtomicReadS8");
1860	BENCH(ASMAtomicReadU16(&s_u16), "ASMAtomicReadU16");
1861	BENCH(ASMAtomicReadS16(&s_i16), "ASMAtomicReadS16");
1862	BENCH(ASMAtomicReadU32(&s_u32), "ASMAtomicReadU32");
1863	BENCH(ASMAtomicReadS32(&s_i32), "ASMAtomicReadS32");
1864	BENCH(ASMAtomicReadU64(&s_u64), "ASMAtomicReadU64");
1865	BENCH(ASMAtomicReadS64(&s_i64), "ASMAtomicReadS64");
1866	BENCH(ASMAtomicUoWriteU8(&s_u8, 0), "ASMAtomicUoWriteU8");
1867	BENCH(ASMAtomicUoWriteS8(&s_i8, 0), "ASMAtomicUoWriteS8");
1868	BENCH(ASMAtomicUoWriteU16(&s_u16, 0), "ASMAtomicUoWriteU16");
1869	BENCH(ASMAtomicUoWriteS16(&s_i16, 0), "ASMAtomicUoWriteS16");
1870	BENCH(ASMAtomicUoWriteU32(&s_u32, 0), "ASMAtomicUoWriteU32");
1871	BENCH(ASMAtomicUoWriteS32(&s_i32, 0), "ASMAtomicUoWriteS32");
1872	BENCH(ASMAtomicUoWriteU64(&s_u64, 0), "ASMAtomicUoWriteU64");
1873	BENCH(ASMAtomicUoWriteS64(&s_i64, 0), "ASMAtomicUoWriteS64");
1874	BENCH(ASMAtomicWriteU8(&s_u8, 0), "ASMAtomicWriteU8");
1875	BENCH(ASMAtomicWriteS8(&s_i8, 0), "ASMAtomicWriteS8");
1876	BENCH(ASMAtomicWriteU16(&s_u16, 0), "ASMAtomicWriteU16");
1877	BENCH(ASMAtomicWriteS16(&s_i16, 0), "ASMAtomicWriteS16");
1878	BENCH(ASMAtomicWriteU32(&s_u32, 0), "ASMAtomicWriteU32");
1879	BENCH(ASMAtomicWriteS32(&s_i32, 0), "ASMAtomicWriteS32");
1880	BENCH(ASMAtomicWriteU64(&s_u64, 0), "ASMAtomicWriteU64");
1881	BENCH(ASMAtomicWriteS64(&s_i64, 0), "ASMAtomicWriteS64");
1882	BENCH(ASMAtomicXchgU8(&s_u8, 0), "ASMAtomicXchgU8");
1883	BENCH(ASMAtomicXchgS8(&s_i8, 0), "ASMAtomicXchgS8");
1884	BENCH(ASMAtomicXchgU16(&s_u16, 0), "ASMAtomicXchgU16");
1885	BENCH(ASMAtomicXchgS16(&s_i16, 0), "ASMAtomicXchgS16");
1886	BENCH(ASMAtomicXchgU32(&s_u32, 0), "ASMAtomicXchgU32");
1887	BENCH(ASMAtomicXchgS32(&s_i32, 0), "ASMAtomicXchgS32");
1888	BENCH(ASMAtomicXchgU64(&s_u64, 0), "ASMAtomicXchgU64");
1889	BENCH(ASMAtomicXchgS64(&s_i64, 0), "ASMAtomicXchgS64");
1890	BENCH(ASMAtomicCmpXchgU32(&s_u32, 0, 0), "ASMAtomicCmpXchgU32");
1891	BENCH(ASMAtomicCmpXchgS32(&s_i32, 0, 0), "ASMAtomicCmpXchgS32");
1892	BENCH(ASMAtomicCmpXchgU64(&s_u64, 0, 0), "ASMAtomicCmpXchgU64");
1893	BENCH(ASMAtomicCmpXchgS64(&s_i64, 0, 0), "ASMAtomicCmpXchgS64");
1894	BENCH(ASMAtomicCmpXchgU32(&s_u32, 0, 1), "ASMAtomicCmpXchgU32/neg");
1895	BENCH(ASMAtomicCmpXchgS32(&s_i32, 0, 1), "ASMAtomicCmpXchgS32/neg");
1896	BENCH(ASMAtomicCmpXchgU64(&s_u64, 0, 1), "ASMAtomicCmpXchgU64/neg");
1897	BENCH(ASMAtomicCmpXchgS64(&s_i64, 0, 1), "ASMAtomicCmpXchgS64/neg");
1898	BENCH(ASMAtomicIncU32(&s_u32), "ASMAtomicIncU32");
1899	BENCH(ASMAtomicIncS32(&s_i32), "ASMAtomicIncS32");
1900	BENCH(ASMAtomicDecU32(&s_u32), "ASMAtomicDecU32");
1901	BENCH(ASMAtomicDecS32(&s_i32), "ASMAtomicDecS32");
1902	BENCH(ASMAtomicAddU32(&s_u32, 5), "ASMAtomicAddU32");
1903	BENCH(ASMAtomicAddS32(&s_i32, 5), "ASMAtomicAddS32");
1904	BENCH(ASMAtomicUoIncU32(&s_u32), "ASMAtomicUoIncU32");
1905	BENCH(ASMAtomicUoDecU32(&s_u32), "ASMAtomicUoDecU32");
1906	BENCH(ASMAtomicUoAndU32(&s_u32, 0xffffffff), "ASMAtomicUoAndU32");
1907	BENCH(ASMAtomicUoOrU32(&s_u32, 0xffffffff), "ASMAtomicUoOrU32");
1908	#if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1909	BENCH_TSC(ASMSerializeInstructionCpuId(), "ASMSerializeInstructionCpuId");
1910	BENCH_TSC(ASMSerializeInstructionIRet(), "ASMSerializeInstructionIRet");
1911	#endif
1912
1913	/* The Darwin gcc does not like this ... */
1914	#if !defined(RT_OS_DARWIN) && !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
1915	BENCH(s_u8 = ASMGetApicId(), "ASMGetApicId");
1916	BENCH(s_u32 = ASMGetApicIdExt0B(), "ASMGetApicIdExt0B");
1917	BENCH(s_u32 = ASMGetApicIdExt8000001E(), "ASMGetApicIdExt8000001E");
1918	#endif
1919	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
1920	uint32_t uAux;
1921	if ( ASMHasCpuId()
1922	&& ASMIsValidExtRange(ASMCpuId_EAX(0x80000000))
1923	&& (ASMCpuId_EDX(0x80000001) & X86_CPUID_EXT_FEATURE_EDX_RDTSCP) )
1924	{
1925	BENCH_TSC(ASMSerializeInstructionRdTscp(), "ASMSerializeInstructionRdTscp");
1926	BENCH(s_u64 = ASMReadTscWithAux(&uAux), "ASMReadTscWithAux");
1927	}
1928	BENCH(s_u64 = ASMReadTSC(), "ASMReadTSC");
1929	union
1930	{
1931	uint64_t u64[2];
1932	RTIDTR Unaligned;
1933	struct
1934	{
1935	uint16_t abPadding[3];
1936	RTIDTR Aligned;
1937	} s;
1938	} uBuf;
1939	Assert(((uintptr_t)&uBuf.Unaligned.pIdt & (sizeof(uintptr_t) - 1)) != 0);
1940	BENCH(ASMGetIDTR(&uBuf.Unaligned), "ASMGetIDTR/unaligned");
1941	Assert(((uintptr_t)&uBuf.s.Aligned.pIdt & (sizeof(uintptr_t) - 1)) == 0);
1942	BENCH(ASMGetIDTR(&uBuf.s.Aligned), "ASMGetIDTR/aligned");
1943	#endif
1944
1945	#undef BENCH
1946	}
1947
1948
1949	int main(int argc, char **argv)
1950	{
1951	RT_NOREF_PV(argc); RT_NOREF_PV(argv);
1952
1953	int rc = RTTestInitAndCreate("tstRTInlineAsm", &g_hTest);
1954	if (rc)
1955	return rc;
1956	RTTestBanner(g_hTest);
1957
1958	/*
1959	* Execute the tests.
1960	*/
1961	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
1962	tstASMCpuId();
1963	//bruteForceCpuId();
1964	#endif
1965	#if 1
1966	tstASMAtomicXchgU8();
1967	tstASMAtomicXchgU16();
1968	tstASMAtomicXchgU32();
1969	tstASMAtomicXchgU64();
1970	tstASMAtomicXchgPtr();
1971	tstASMAtomicCmpXchgU8();
1972	tstASMAtomicCmpXchgU32();
1973	tstASMAtomicCmpXchgU64();
1974	tstASMAtomicCmpXchgExU32();
1975	tstASMAtomicCmpXchgExU64();
1976	tstASMAtomicReadU64();
1977	tstASMAtomicUoReadU64();
1978
1979	tstASMAtomicAddS32();
1980	tstASMAtomicAddS64();
1981	tstASMAtomicDecIncS32();
1982	tstASMAtomicDecIncS64();
1983	tstASMAtomicAndOrU32();
1984	tstASMAtomicAndOrU64();
1985
1986	tstASMAtomicUoIncU32();
1987	tstASMAtomicUoDecU32();
1988	tstASMAtomicUoAndOrU32();
1989
1990	tstASMMemZeroPage();
1991	tstASMMemIsZeroPage(g_hTest);
1992	tstASMMemFirstMismatchingU8(g_hTest);
1993	tstASMMemZero32();
1994	tstASMMemFill32();
1995
1996	tstASMMath();
1997
1998	tstASMByteSwap();
1999
2000	tstASMBench();
2001	#endif
2002
2003	/*
2004	* Show the result.
2005	*/
2006	return RTTestSummaryAndDestroy(g_hTest);
2007	}
2008

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

source: vbox/trunk/src/VBox/Runtime/testcase/tstRTInlineAsm.cpp@ 85546

Download in other formats: