tstRTInlineAsm.cpp@ 61311

Last change on this file since 61311 was 61311, checked in by vboxsync, 9 years ago
tstRTInlineAsm: fixes for leaf EAX=7
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 73.3 KB

Line
1	/* $Id: tstRTInlineAsm.cpp 61311 2016-05-30 16:45:00Z vboxsync $ */
2	/** @file
3	* IPRT Testcase - inline assembly.
4	*/
5
6	/*
7	* Copyright (C) 2006-2015 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	* Done testing, dump the information.
209	*/
210	RTTestIPrintf(RTTESTLVL_ALWAYS, "CPUID Dump\n");
211	ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
212	const uint32_t cFunctions = s.uEAX;
213
214	/* raw dump */
215	RTTestIPrintf(RTTESTLVL_ALWAYS,
216	"\n"
217	" RAW Standard CPUIDs\n"
218	"Function eax ebx ecx edx\n");
219	for (unsigned iStd = 0; iStd <= cFunctions + 3; iStd++)
220	{
221	ASMCpuId_Idx_ECX(iStd, 0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
222	RTTestIPrintf(RTTESTLVL_ALWAYS, "%08x %08x %08x %08x %08x%s\n",
223	iStd, s.uEAX, s.uEBX, s.uECX, s.uEDX, iStd <= cFunctions ? "" : "*");
224
225	/* Leaf 04 and leaf 0d output depend on the initial value of ECX
226	* The same seems to apply to invalid standard functions */
227	if (iStd > cFunctions)
228	continue;
229	if (iStd != 0x04 && iStd != 0x07 && iStd != 0x0b && iStd != 0x0d)
230	{
231	u32 = ASMCpuId_EAX(iStd);
232	CHECKVAL(u32, s.uEAX, "%x");
233
234	uint32_t u32EbxMask = UINT32_MAX;
235	if (iStd == 1)
236	u32EbxMask = UINT32_C(0x00ffffff); /* Omit the local apic ID in case we're rescheduled. */
237	u32 = ASMCpuId_EBX(iStd);
238	CHECKVAL(u32 & u32EbxMask, s.uEBX & u32EbxMask, "%x");
239
240	u32 = ASMCpuId_ECX(iStd);
241	CHECKVAL(u32, s.uECX, "%x");
242	u32 = ASMCpuId_EDX(iStd);
243	CHECKVAL(u32, s.uEDX, "%x");
244
245	uECX2 = s.uECX - 1;
246	uEDX2 = s.uEDX - 1;
247	ASMCpuId_ECX_EDX(iStd, &uECX2, &uEDX2);
248	CHECKVAL(uECX2, s.uECX, "%x");
249	CHECKVAL(uEDX2, s.uEDX, "%x");
250	}
251
252	if (iStd == 0x04)
253	for (uint32_t uECX = 1; s.uEAX & 0x1f; uECX++)
254	{
255	ASMCpuId_Idx_ECX(iStd, uECX, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
256	RTTestIPrintf(RTTESTLVL_ALWAYS, " [%02x] %08x %08x %08x %08x\n", uECX, s.uEAX, s.uEBX, s.uECX, s.uEDX);
257	RTTESTI_CHECK_BREAK(uECX < 128);
258	}
259	else if (iStd == 0x07)
260	{
261	uint32_t uMax = s.uEAX;
262	for (uint32_t uECX = 1; uECX < uMax; 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	}
269	else if (iStd == 0x0b)
270	for (uint32_t uECX = 1; (s.uEAX & 0x1f) && (s.uEBX & 0xffff); uECX++)
271	{
272	ASMCpuId_Idx_ECX(iStd, uECX, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
273	RTTestIPrintf(RTTESTLVL_ALWAYS, " [%02x] %08x %08x %08x %08x\n", uECX, s.uEAX, s.uEBX, s.uECX, s.uEDX);
274	RTTESTI_CHECK_BREAK(uECX < 128);
275	}
276	else if (iStd == 0x0d)
277	for (uint32_t uECX = 1; s.uEAX != 0 \|\| s.uEBX != 0 \|\| s.uECX != 0 \|\| s.uEDX != 0; uECX++)
278	{
279	ASMCpuId_Idx_ECX(iStd, uECX, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
280	RTTestIPrintf(RTTESTLVL_ALWAYS, " [%02x] %08x %08x %08x %08x\n", uECX, s.uEAX, s.uEBX, s.uECX, s.uEDX);
281	RTTESTI_CHECK_BREAK(uECX < 128);
282	}
283	}
284
285	/*
286	* Understandable output
287	*/
288	ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
289	RTTestIPrintf(RTTESTLVL_ALWAYS,
290	"Name: %.04s%.04s%.04s\n"
291	"Support: 0-%u\n",
292	&s.uEBX, &s.uEDX, &s.uECX, s.uEAX);
293	bool const fIntel = ASMIsIntelCpuEx(s.uEBX, s.uECX, s.uEDX);
294
295	/*
296	* Get Features.
297	*/
298	if (cFunctions >= 1)
299	{
300	static const char * const s_apszTypes[4] = { "primary", "overdrive", "MP", "reserved" };
301	ASMCpuId(1, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
302	RTTestIPrintf(RTTESTLVL_ALWAYS,
303	"Family: %#x \tExtended: %#x \tEffective: %#x\n"
304	"Model: %#x \tExtended: %#x \tEffective: %#x\n"
305	"Stepping: %d\n"
306	"Type: %d (%s)\n"
307	"APIC ID: %#04x\n"
308	"Logical CPUs: %d\n"
309	"CLFLUSH Size: %d\n"
310	"Brand ID: %#04x\n",
311	(s.uEAX >> 8) & 0xf, (s.uEAX >> 20) & 0x7f, ASMGetCpuFamily(s.uEAX),
312	(s.uEAX >> 4) & 0xf, (s.uEAX >> 16) & 0x0f, ASMGetCpuModel(s.uEAX, fIntel),
313	ASMGetCpuStepping(s.uEAX),
314	(s.uEAX >> 12) & 0x3, s_apszTypes[(s.uEAX >> 12) & 0x3],
315	(s.uEBX >> 24) & 0xff,
316	(s.uEBX >> 16) & 0xff,
317	(s.uEBX >> 8) & 0xff,
318	(s.uEBX >> 0) & 0xff);
319
320	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features EDX: ");
321	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FPU");
322	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VME");
323	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DE");
324	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE");
325	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TSC");
326	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MSR");
327	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAE");
328	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCE");
329	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CX8");
330	if (s.uEDX & RT_BIT(9)) RTTestIPrintf(RTTESTLVL_ALWAYS, " APIC");
331	if (s.uEDX & RT_BIT(10)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 10");
332	if (s.uEDX & RT_BIT(11)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SEP");
333	if (s.uEDX & RT_BIT(12)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MTRR");
334	if (s.uEDX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PGE");
335	if (s.uEDX & RT_BIT(14)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCA");
336	if (s.uEDX & RT_BIT(15)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMOV");
337	if (s.uEDX & RT_BIT(16)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAT");
338	if (s.uEDX & RT_BIT(17)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE36");
339	if (s.uEDX & RT_BIT(18)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSN");
340	if (s.uEDX & RT_BIT(19)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CLFSH");
341	if (s.uEDX & RT_BIT(20)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 20");
342	if (s.uEDX & RT_BIT(21)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DS");
343	if (s.uEDX & RT_BIT(22)) RTTestIPrintf(RTTESTLVL_ALWAYS, " ACPI");
344	if (s.uEDX & RT_BIT(23)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MMX");
345	if (s.uEDX & RT_BIT(24)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FXSR");
346	if (s.uEDX & RT_BIT(25)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE");
347	if (s.uEDX & RT_BIT(26)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE2");
348	if (s.uEDX & RT_BIT(27)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SS");
349	if (s.uEDX & RT_BIT(28)) RTTestIPrintf(RTTESTLVL_ALWAYS, " HTT");
350	if (s.uEDX & RT_BIT(29)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 29");
351	if (s.uEDX & RT_BIT(30)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 30");
352	if (s.uEDX & RT_BIT(31)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 31");
353	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
354
355	/** @todo check intel docs. */
356	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features ECX: ");
357	if (s.uECX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE3");
358	for (iBit = 1; iBit < 13; iBit++)
359	if (s.uECX & RT_BIT(iBit))
360	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
361	if (s.uECX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CX16");
362	for (iBit = 14; iBit < 32; iBit++)
363	if (s.uECX & RT_BIT(iBit))
364	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
365	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
366	}
367
368	/*
369	* Extended.
370	* Implemented after AMD specs.
371	*/
372	/** @todo check out the intel specs. */
373	ASMCpuId(0x80000000, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
374	if (!s.uEAX && !s.uEBX && !s.uECX && !s.uEDX)
375	{
376	RTTestIPrintf(RTTESTLVL_ALWAYS, "No extended CPUID info? Check the manual on how to detect this...\n");
377	return;
378	}
379	const uint32_t cExtFunctions = s.uEAX \| 0x80000000;
380
381	/* raw dump */
382	RTTestIPrintf(RTTESTLVL_ALWAYS,
383	"\n"
384	" RAW Extended CPUIDs\n"
385	"Function eax ebx ecx edx\n");
386	for (unsigned iExt = 0x80000000; iExt <= cExtFunctions + 3; iExt++)
387	{
388	ASMCpuId(iExt, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
389	RTTestIPrintf(RTTESTLVL_ALWAYS, "%08x %08x %08x %08x %08x%s\n",
390	iExt, s.uEAX, s.uEBX, s.uECX, s.uEDX, iExt <= cExtFunctions ? "" : "*");
391
392	if (iExt > cExtFunctions)
393	continue; /* Invalid extended functions seems change the value if ECX changes */
394	if (iExt == 0x8000001d)
395	continue; /* Takes cache level in ecx. */
396
397	u32 = ASMCpuId_EAX(iExt);
398	CHECKVAL(u32, s.uEAX, "%x");
399	u32 = ASMCpuId_EBX(iExt);
400	CHECKVAL(u32, s.uEBX, "%x");
401	u32 = ASMCpuId_ECX(iExt);
402	CHECKVAL(u32, s.uECX, "%x");
403	u32 = ASMCpuId_EDX(iExt);
404	CHECKVAL(u32, s.uEDX, "%x");
405
406	uECX2 = s.uECX - 1;
407	uEDX2 = s.uEDX - 1;
408	ASMCpuId_ECX_EDX(iExt, &uECX2, &uEDX2);
409	CHECKVAL(uECX2, s.uECX, "%x");
410	CHECKVAL(uEDX2, s.uEDX, "%x");
411	}
412
413	/*
414	* Understandable output
415	*/
416	ASMCpuId(0x80000000, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
417	RTTestIPrintf(RTTESTLVL_ALWAYS,
418	"Ext Name: %.4s%.4s%.4s\n"
419	"Ext Supports: 0x80000000-%#010x\n",
420	&s.uEBX, &s.uEDX, &s.uECX, s.uEAX);
421
422	if (cExtFunctions >= 0x80000001)
423	{
424	ASMCpuId(0x80000001, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
425	RTTestIPrintf(RTTESTLVL_ALWAYS,
426	"Family: %#x \tExtended: %#x \tEffective: %#x\n"
427	"Model: %#x \tExtended: %#x \tEffective: %#x\n"
428	"Stepping: %d\n"
429	"Brand ID: %#05x\n",
430	(s.uEAX >> 8) & 0xf, (s.uEAX >> 20) & 0x7f, ASMGetCpuFamily(s.uEAX),
431	(s.uEAX >> 4) & 0xf, (s.uEAX >> 16) & 0x0f, ASMGetCpuModel(s.uEAX, fIntel),
432	ASMGetCpuStepping(s.uEAX),
433	s.uEBX & 0xfff);
434
435	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features EDX: ");
436	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FPU");
437	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VME");
438	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DE");
439	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE");
440	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TSC");
441	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MSR");
442	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAE");
443	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCE");
444	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMPXCHG8B");
445	if (s.uEDX & RT_BIT(9)) RTTestIPrintf(RTTESTLVL_ALWAYS, " APIC");
446	if (s.uEDX & RT_BIT(10)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 10");
447	if (s.uEDX & RT_BIT(11)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SysCallSysRet");
448	if (s.uEDX & RT_BIT(12)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MTRR");
449	if (s.uEDX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PGE");
450	if (s.uEDX & RT_BIT(14)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCA");
451	if (s.uEDX & RT_BIT(15)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMOV");
452	if (s.uEDX & RT_BIT(16)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAT");
453	if (s.uEDX & RT_BIT(17)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE36");
454	if (s.uEDX & RT_BIT(18)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 18");
455	if (s.uEDX & RT_BIT(19)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 19");
456	if (s.uEDX & RT_BIT(20)) RTTestIPrintf(RTTESTLVL_ALWAYS, " NX");
457	if (s.uEDX & RT_BIT(21)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 21");
458	if (s.uEDX & RT_BIT(22)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MmxExt");
459	if (s.uEDX & RT_BIT(23)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MMX");
460	if (s.uEDX & RT_BIT(24)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FXSR");
461	if (s.uEDX & RT_BIT(25)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FastFXSR");
462	if (s.uEDX & RT_BIT(26)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 26");
463	if (s.uEDX & RT_BIT(27)) RTTestIPrintf(RTTESTLVL_ALWAYS, " RDTSCP");
464	if (s.uEDX & RT_BIT(28)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 28");
465	if (s.uEDX & RT_BIT(29)) RTTestIPrintf(RTTESTLVL_ALWAYS, " LongMode");
466	if (s.uEDX & RT_BIT(30)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3DNowExt");
467	if (s.uEDX & RT_BIT(31)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3DNow");
468	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
469
470	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features ECX: ");
471	if (s.uECX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " LahfSahf");
472	if (s.uECX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CmpLegacy");
473	if (s.uECX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SVM");
474	if (s.uECX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3");
475	if (s.uECX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " AltMovCr8");
476	for (iBit = 5; iBit < 32; iBit++)
477	if (s.uECX & RT_BIT(iBit))
478	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
479	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
480	}
481
482	char szString[443+1] = {0};
483	if (cExtFunctions >= 0x80000002)
484	ASMCpuId(0x80000002, &szString[0 + 0], &szString[0 + 4], &szString[0 + 8], &szString[0 + 12]);
485	if (cExtFunctions >= 0x80000003)
486	ASMCpuId(0x80000003, &szString[16 + 0], &szString[16 + 4], &szString[16 + 8], &szString[16 + 12]);
487	if (cExtFunctions >= 0x80000004)
488	ASMCpuId(0x80000004, &szString[32 + 0], &szString[32 + 4], &szString[32 + 8], &szString[32 + 12]);
489	if (cExtFunctions >= 0x80000002)
490	RTTestIPrintf(RTTESTLVL_ALWAYS, "Full Name: %s\n", szString);
491
492	if (cExtFunctions >= 0x80000005)
493	{
494	ASMCpuId(0x80000005, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
495	RTTestIPrintf(RTTESTLVL_ALWAYS,
496	"TLB 2/4M Instr/Uni: %s %3d entries\n"
497	"TLB 2/4M Data: %s %3d entries\n",
498	getCacheAss((s.uEAX >> 8) & 0xff), (s.uEAX >> 0) & 0xff,
499	getCacheAss((s.uEAX >> 24) & 0xff), (s.uEAX >> 16) & 0xff);
500	RTTestIPrintf(RTTESTLVL_ALWAYS,
501	"TLB 4K Instr/Uni: %s %3d entries\n"
502	"TLB 4K Data: %s %3d entries\n",
503	getCacheAss((s.uEBX >> 8) & 0xff), (s.uEBX >> 0) & 0xff,
504	getCacheAss((s.uEBX >> 24) & 0xff), (s.uEBX >> 16) & 0xff);
505	RTTestIPrintf(RTTESTLVL_ALWAYS,
506	"L1 Instr Cache Line Size: %d bytes\n"
507	"L1 Instr Cache Lines Per Tag: %d\n"
508	"L1 Instr Cache Associativity: %s\n"
509	"L1 Instr Cache Size: %d KB\n",
510	(s.uEDX >> 0) & 0xff,
511	(s.uEDX >> 8) & 0xff,
512	getCacheAss((s.uEDX >> 16) & 0xff),
513	(s.uEDX >> 24) & 0xff);
514	RTTestIPrintf(RTTESTLVL_ALWAYS,
515	"L1 Data Cache Line Size: %d bytes\n"
516	"L1 Data Cache Lines Per Tag: %d\n"
517	"L1 Data Cache Associativity: %s\n"
518	"L1 Data Cache Size: %d KB\n",
519	(s.uECX >> 0) & 0xff,
520	(s.uECX >> 8) & 0xff,
521	getCacheAss((s.uECX >> 16) & 0xff),
522	(s.uECX >> 24) & 0xff);
523	}
524
525	if (cExtFunctions >= 0x80000006)
526	{
527	ASMCpuId(0x80000006, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
528	RTTestIPrintf(RTTESTLVL_ALWAYS,
529	"L2 TLB 2/4M Instr/Uni: %s %4d entries\n"
530	"L2 TLB 2/4M Data: %s %4d entries\n",
531	getL2CacheAss((s.uEAX >> 12) & 0xf), (s.uEAX >> 0) & 0xfff,
532	getL2CacheAss((s.uEAX >> 28) & 0xf), (s.uEAX >> 16) & 0xfff);
533	RTTestIPrintf(RTTESTLVL_ALWAYS,
534	"L2 TLB 4K Instr/Uni: %s %4d entries\n"
535	"L2 TLB 4K Data: %s %4d entries\n",
536	getL2CacheAss((s.uEBX >> 12) & 0xf), (s.uEBX >> 0) & 0xfff,
537	getL2CacheAss((s.uEBX >> 28) & 0xf), (s.uEBX >> 16) & 0xfff);
538	RTTestIPrintf(RTTESTLVL_ALWAYS,
539	"L2 Cache Line Size: %d bytes\n"
540	"L2 Cache Lines Per Tag: %d\n"
541	"L2 Cache Associativity: %s\n"
542	"L2 Cache Size: %d KB\n",
543	(s.uEDX >> 0) & 0xff,
544	(s.uEDX >> 8) & 0xf,
545	getL2CacheAss((s.uEDX >> 12) & 0xf),
546	(s.uEDX >> 16) & 0xffff);
547	}
548
549	if (cExtFunctions >= 0x80000007)
550	{
551	ASMCpuId(0x80000007, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
552	RTTestIPrintf(RTTESTLVL_ALWAYS, "APM Features: ");
553	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TS");
554	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FID");
555	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VID");
556	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TTP");
557	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TM");
558	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " STC");
559	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 6");
560	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 7");
561	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TscInvariant");
562	for (iBit = 9; iBit < 32; iBit++)
563	if (s.uEDX & RT_BIT(iBit))
564	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
565	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
566	}
567
568	if (cExtFunctions >= 0x80000008)
569	{
570	ASMCpuId(0x80000008, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
571	RTTestIPrintf(RTTESTLVL_ALWAYS,
572	"Physical Address Width: %d bits\n"
573	"Virtual Address Width: %d bits\n"
574	"Guest Physical Address Width: %d bits\n",
575	(s.uEAX >> 0) & 0xff,
576	(s.uEAX >> 8) & 0xff,
577	(s.uEAX >> 16) & 0xff);
578	RTTestIPrintf(RTTESTLVL_ALWAYS,
579	"Physical Core Count: %d\n",
580	((s.uECX >> 0) & 0xff) + 1);
581	if ((s.uECX >> 12) & 0xf)
582	RTTestIPrintf(RTTESTLVL_ALWAYS, "ApicIdCoreIdSize: %d bits\n", (s.uECX >> 12) & 0xf);
583	}
584
585	if (cExtFunctions >= 0x8000000a)
586	{
587	ASMCpuId(0x8000000a, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
588	RTTestIPrintf(RTTESTLVL_ALWAYS,
589	"SVM Revision: %d (%#x)\n"
590	"Number of Address Space IDs: %d (%#x)\n",
591	s.uEAX & 0xff, s.uEAX & 0xff,
592	s.uEBX, s.uEBX);
593	}
594	}
595
596	# if 0
597	static void bruteForceCpuId(void)
598	{
599	RTTestISub("brute force CPUID leafs");
600	uint32_t auPrevValues[4] = { 0, 0, 0, 0};
601	uint32_t uLeaf = 0;
602	do
603	{
604	uint32_t auValues[4];
605	ASMCpuIdExSlow(uLeaf, 0, 0, 0, &auValues[0], &auValues[1], &auValues[2], &auValues[3]);
606	if ( (auValues[0] != auPrevValues[0] && auValues[0] != uLeaf)
607	\|\| (auValues[1] != auPrevValues[1] && auValues[1] != 0)
608	\|\| (auValues[2] != auPrevValues[2] && auValues[2] != 0)
609	\|\| (auValues[3] != auPrevValues[3] && auValues[3] != 0)
610	\|\| (uLeaf & (UINT32_C(0x08000000) - UINT32_C(1))) == 0)
611	{
612	RTTestIPrintf(RTTESTLVL_ALWAYS,
613	"%08x: %08x %08x %08x %08x\n", uLeaf,
614	auValues[0], auValues[1], auValues[2], auValues[3]);
615	}
616	auPrevValues[0] = auValues[0];
617	auPrevValues[1] = auValues[1];
618	auPrevValues[2] = auValues[2];
619	auPrevValues[3] = auValues[3];
620
621	//uint32_t uSubLeaf = 0;
622	//do
623	//{
624	//
625	//
626	//} while (false);
627	} while (uLeaf++ < UINT32_MAX);
628	}
629	# endif
630
631	#endif /* AMD64 \|\| X86 */
632
633	DECLINLINE(void) tstASMAtomicXchgU8Worker(uint8_t volatile *pu8)
634	{
635	*pu8 = 0;
636	CHECKOP(ASMAtomicXchgU8(pu8, 1), 0, "%#x", uint8_t);
637	CHECKVAL(*pu8, 1, "%#x");
638
639	CHECKOP(ASMAtomicXchgU8(pu8, 0), 1, "%#x", uint8_t);
640	CHECKVAL(*pu8, 0, "%#x");
641
642	CHECKOP(ASMAtomicXchgU8(pu8, 0xff), 0, "%#x", uint8_t);
643	CHECKVAL(*pu8, 0xff, "%#x");
644
645	CHECKOP(ASMAtomicXchgU8(pu8, 0x87), 0xffff, "%#x", uint8_t);
646	CHECKVAL(*pu8, 0x87, "%#x");
647	}
648
649
650	static void tstASMAtomicXchgU8(void)
651	{
652	DO_SIMPLE_TEST(ASMAtomicXchgU8, uint8_t);
653	}
654
655
656	DECLINLINE(void) tstASMAtomicXchgU16Worker(uint16_t volatile *pu16)
657	{
658	*pu16 = 0;
659
660	CHECKOP(ASMAtomicXchgU16(pu16, 1), 0, "%#x", uint16_t);
661	CHECKVAL(*pu16, 1, "%#x");
662
663	CHECKOP(ASMAtomicXchgU16(pu16, 0), 1, "%#x", uint16_t);
664	CHECKVAL(*pu16, 0, "%#x");
665
666	CHECKOP(ASMAtomicXchgU16(pu16, 0xffff), 0, "%#x", uint16_t);
667	CHECKVAL(*pu16, 0xffff, "%#x");
668
669	CHECKOP(ASMAtomicXchgU16(pu16, 0x8765), 0xffff, "%#x", uint16_t);
670	CHECKVAL(*pu16, 0x8765, "%#x");
671	}
672
673
674	static void tstASMAtomicXchgU16(void)
675	{
676	DO_SIMPLE_TEST(ASMAtomicXchgU16, uint16_t);
677	}
678
679
680	DECLINLINE(void) tstASMAtomicXchgU32Worker(uint32_t volatile *pu32)
681	{
682	*pu32 = 0;
683
684	CHECKOP(ASMAtomicXchgU32(pu32, 1), 0, "%#x", uint32_t);
685	CHECKVAL(*pu32, 1, "%#x");
686
687	CHECKOP(ASMAtomicXchgU32(pu32, 0), 1, "%#x", uint32_t);
688	CHECKVAL(*pu32, 0, "%#x");
689
690	CHECKOP(ASMAtomicXchgU32(pu32, ~UINT32_C(0)), 0, "%#x", uint32_t);
691	CHECKVAL(*pu32, ~UINT32_C(0), "%#x");
692
693	CHECKOP(ASMAtomicXchgU32(pu32, 0x87654321), ~UINT32_C(0), "%#x", uint32_t);
694	CHECKVAL(*pu32, 0x87654321, "%#x");
695	}
696
697
698	static void tstASMAtomicXchgU32(void)
699	{
700	DO_SIMPLE_TEST(ASMAtomicXchgU32, uint32_t);
701	}
702
703
704	DECLINLINE(void) tstASMAtomicXchgU64Worker(uint64_t volatile *pu64)
705	{
706	*pu64 = 0;
707
708	CHECKOP(ASMAtomicXchgU64(pu64, 1), UINT64_C(0), "%#llx", uint64_t);
709	CHECKVAL(*pu64, UINT64_C(1), "%#llx");
710
711	CHECKOP(ASMAtomicXchgU64(pu64, 0), UINT64_C(1), "%#llx", uint64_t);
712	CHECKVAL(*pu64, UINT64_C(0), "%#llx");
713
714	CHECKOP(ASMAtomicXchgU64(pu64, ~UINT64_C(0)), UINT64_C(0), "%#llx", uint64_t);
715	CHECKVAL(*pu64, ~UINT64_C(0), "%#llx");
716
717	CHECKOP(ASMAtomicXchgU64(pu64, UINT64_C(0xfedcba0987654321)), ~UINT64_C(0), "%#llx", uint64_t);
718	CHECKVAL(*pu64, UINT64_C(0xfedcba0987654321), "%#llx");
719	}
720
721
722	static void tstASMAtomicXchgU64(void)
723	{
724	DO_SIMPLE_TEST(ASMAtomicXchgU64, uint64_t);
725	}
726
727
728	DECLINLINE(void) tstASMAtomicXchgPtrWorker(void * volatile *ppv)
729	{
730	*ppv = NULL;
731
732	CHECKOP(ASMAtomicXchgPtr(ppv, (void )(~(uintptr_t)0)), NULL, "%p", void );
733	CHECKVAL(ppv, (void )(~(uintptr_t)0), "%p");
734
735	CHECKOP(ASMAtomicXchgPtr(ppv, (void )0x87654321), (void )(~(uintptr_t)0), "%p", void *);
736	CHECKVAL(ppv, (void )0x87654321, "%p");
737
738	CHECKOP(ASMAtomicXchgPtr(ppv, NULL), (void )0x87654321, "%p", void );
739	CHECKVAL(*ppv, NULL, "%p");
740	}
741
742
743	static void tstASMAtomicXchgPtr(void)
744	{
745	DO_SIMPLE_TEST(ASMAtomicXchgPtr, void *);
746	}
747
748
749	DECLINLINE(void) tstASMAtomicCmpXchgU8Worker(uint8_t volatile *pu8)
750	{
751	*pu8 = 0xff;
752
753	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0, 0), false, "%d", bool);
754	CHECKVAL(*pu8, 0xff, "%x");
755
756	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0, 0xff), true, "%d", bool);
757	CHECKVAL(*pu8, 0, "%x");
758
759	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0x79, 0xff), false, "%d", bool);
760	CHECKVAL(*pu8, 0, "%x");
761
762	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0x97, 0), true, "%d", bool);
763	CHECKVAL(*pu8, 0x97, "%x");
764	}
765
766
767	static void tstASMAtomicCmpXchgU8(void)
768	{
769	DO_SIMPLE_TEST(ASMAtomicCmpXchgU8, uint8_t);
770	}
771
772
773	DECLINLINE(void) tstASMAtomicCmpXchgU32Worker(uint32_t volatile *pu32)
774	{
775	*pu32 = UINT32_C(0xffffffff);
776
777	CHECKOP(ASMAtomicCmpXchgU32(pu32, 0, 0), false, "%d", bool);
778	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
779
780	CHECKOP(ASMAtomicCmpXchgU32(pu32, 0, UINT32_C(0xffffffff)), true, "%d", bool);
781	CHECKVAL(*pu32, 0, "%x");
782
783	CHECKOP(ASMAtomicCmpXchgU32(pu32, UINT32_C(0x8008efd), UINT32_C(0xffffffff)), false, "%d", bool);
784	CHECKVAL(*pu32, 0, "%x");
785
786	CHECKOP(ASMAtomicCmpXchgU32(pu32, UINT32_C(0x8008efd), 0), true, "%d", bool);
787	CHECKVAL(*pu32, UINT32_C(0x8008efd), "%x");
788	}
789
790
791	static void tstASMAtomicCmpXchgU32(void)
792	{
793	DO_SIMPLE_TEST(ASMAtomicCmpXchgU32, uint32_t);
794	}
795
796
797
798	DECLINLINE(void) tstASMAtomicCmpXchgU64Worker(uint64_t volatile *pu64)
799	{
800	*pu64 = UINT64_C(0xffffffffffffff);
801
802	CHECKOP(ASMAtomicCmpXchgU64(pu64, 0, 0), false, "%d", bool);
803	CHECKVAL(*pu64, UINT64_C(0xffffffffffffff), "%#llx");
804
805	CHECKOP(ASMAtomicCmpXchgU64(pu64, 0, UINT64_C(0xffffffffffffff)), true, "%d", bool);
806	CHECKVAL(*pu64, 0, "%x");
807
808	CHECKOP(ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), UINT64_C(0xffffffff)), false, "%d", bool);
809	CHECKVAL(*pu64, 0, "%x");
810
811	CHECKOP(ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), UINT64_C(0xffffffff00000000)), false, "%d", bool);
812	CHECKVAL(*pu64, 0, "%x");
813
814	CHECKOP(ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), 0), true, "%d", bool);
815	CHECKVAL(*pu64, UINT64_C(0x80040008008efd), "%#llx");
816	}
817
818
819	static void tstASMAtomicCmpXchgU64(void)
820	{
821	DO_SIMPLE_TEST(ASMAtomicCmpXchgU64, uint64_t);
822	}
823
824
825	DECLINLINE(void) tstASMAtomicCmpXchgExU32Worker(uint32_t volatile *pu32)
826	{
827	*pu32 = UINT32_C(0xffffffff);
828	uint32_t u32Old = UINT32_C(0x80005111);
829
830	CHECKOP(ASMAtomicCmpXchgExU32(pu32, 0, 0, &u32Old), false, "%d", bool);
831	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
832	CHECKVAL(u32Old, UINT32_C(0xffffffff), "%x");
833
834	CHECKOP(ASMAtomicCmpXchgExU32(pu32, 0, UINT32_C(0xffffffff), &u32Old), true, "%d", bool);
835	CHECKVAL(*pu32, 0, "%x");
836	CHECKVAL(u32Old, UINT32_C(0xffffffff), "%x");
837
838	CHECKOP(ASMAtomicCmpXchgExU32(pu32, UINT32_C(0x8008efd), UINT32_C(0xffffffff), &u32Old), false, "%d", bool);
839	CHECKVAL(*pu32, 0, "%x");
840	CHECKVAL(u32Old, 0, "%x");
841
842	CHECKOP(ASMAtomicCmpXchgExU32(pu32, UINT32_C(0x8008efd), 0, &u32Old), true, "%d", bool);
843	CHECKVAL(*pu32, UINT32_C(0x8008efd), "%x");
844	CHECKVAL(u32Old, 0, "%x");
845
846	CHECKOP(ASMAtomicCmpXchgExU32(pu32, 0, UINT32_C(0x8008efd), &u32Old), true, "%d", bool);
847	CHECKVAL(*pu32, 0, "%x");
848	CHECKVAL(u32Old, UINT32_C(0x8008efd), "%x");
849	}
850
851
852	static void tstASMAtomicCmpXchgExU32(void)
853	{
854	DO_SIMPLE_TEST(ASMAtomicCmpXchgExU32, uint32_t);
855	}
856
857
858	DECLINLINE(void) tstASMAtomicCmpXchgExU64Worker(uint64_t volatile *pu64)
859	{
860	*pu64 = UINT64_C(0xffffffffffffffff);
861	uint64_t u64Old = UINT64_C(0x8000000051111111);
862
863	CHECKOP(ASMAtomicCmpXchgExU64(pu64, 0, 0, &u64Old), false, "%d", bool);
864	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%llx");
865	CHECKVAL(u64Old, UINT64_C(0xffffffffffffffff), "%llx");
866
867	CHECKOP(ASMAtomicCmpXchgExU64(pu64, 0, UINT64_C(0xffffffffffffffff), &u64Old), true, "%d", bool);
868	CHECKVAL(*pu64, UINT64_C(0), "%llx");
869	CHECKVAL(u64Old, UINT64_C(0xffffffffffffffff), "%llx");
870
871	CHECKOP(ASMAtomicCmpXchgExU64(pu64, UINT64_C(0x80040008008efd), 0xffffffff, &u64Old), false, "%d", bool);
872	CHECKVAL(*pu64, UINT64_C(0), "%llx");
873	CHECKVAL(u64Old, UINT64_C(0), "%llx");
874
875	CHECKOP(ASMAtomicCmpXchgExU64(pu64, UINT64_C(0x80040008008efd), UINT64_C(0xffffffff00000000), &u64Old), false, "%d", bool);
876	CHECKVAL(*pu64, UINT64_C(0), "%llx");
877	CHECKVAL(u64Old, UINT64_C(0), "%llx");
878
879	CHECKOP(ASMAtomicCmpXchgExU64(pu64, UINT64_C(0x80040008008efd), 0, &u64Old), true, "%d", bool);
880	CHECKVAL(*pu64, UINT64_C(0x80040008008efd), "%llx");
881	CHECKVAL(u64Old, UINT64_C(0), "%llx");
882
883	CHECKOP(ASMAtomicCmpXchgExU64(pu64, 0, UINT64_C(0x80040008008efd), &u64Old), true, "%d", bool);
884	CHECKVAL(*pu64, UINT64_C(0), "%llx");
885	CHECKVAL(u64Old, UINT64_C(0x80040008008efd), "%llx");
886	}
887
888
889	static void tstASMAtomicCmpXchgExU64(void)
890	{
891	DO_SIMPLE_TEST(ASMAtomicCmpXchgExU64, uint64_t);
892	}
893
894
895	DECLINLINE(void) tstASMAtomicReadU64Worker(uint64_t volatile *pu64)
896	{
897	*pu64 = 0;
898
899	CHECKOP(ASMAtomicReadU64(pu64), UINT64_C(0), "%#llx", uint64_t);
900	CHECKVAL(*pu64, UINT64_C(0), "%#llx");
901
902	*pu64 = ~UINT64_C(0);
903	CHECKOP(ASMAtomicReadU64(pu64), ~UINT64_C(0), "%#llx", uint64_t);
904	CHECKVAL(*pu64, ~UINT64_C(0), "%#llx");
905
906	*pu64 = UINT64_C(0xfedcba0987654321);
907	CHECKOP(ASMAtomicReadU64(pu64), UINT64_C(0xfedcba0987654321), "%#llx", uint64_t);
908	CHECKVAL(*pu64, UINT64_C(0xfedcba0987654321), "%#llx");
909	}
910
911
912	static void tstASMAtomicReadU64(void)
913	{
914	DO_SIMPLE_TEST(ASMAtomicReadU64, uint64_t);
915	}
916
917
918	DECLINLINE(void) tstASMAtomicUoReadU64Worker(uint64_t volatile *pu64)
919	{
920	*pu64 = 0;
921
922	CHECKOP(ASMAtomicUoReadU64(pu64), UINT64_C(0), "%#llx", uint64_t);
923	CHECKVAL(*pu64, UINT64_C(0), "%#llx");
924
925	*pu64 = ~UINT64_C(0);
926	CHECKOP(ASMAtomicUoReadU64(pu64), ~UINT64_C(0), "%#llx", uint64_t);
927	CHECKVAL(*pu64, ~UINT64_C(0), "%#llx");
928
929	*pu64 = UINT64_C(0xfedcba0987654321);
930	CHECKOP(ASMAtomicUoReadU64(pu64), UINT64_C(0xfedcba0987654321), "%#llx", uint64_t);
931	CHECKVAL(*pu64, UINT64_C(0xfedcba0987654321), "%#llx");
932	}
933
934
935	static void tstASMAtomicUoReadU64(void)
936	{
937	DO_SIMPLE_TEST(ASMAtomicUoReadU64, uint64_t);
938	}
939
940
941	DECLINLINE(void) tstASMAtomicAddS32Worker(int32_t *pi32)
942	{
943	int32_t i32Rc;
944	*pi32 = 10;
945	#define MYCHECK(op, rc, val) \
946	do { \
947	i32Rc = op; \
948	if (i32Rc != (rc)) \
949	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %d expected %d\n", __FUNCTION__, __LINE__, #op, i32Rc, rc); \
950	if (*pi32 != (val)) \
951	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi32=%d expected %d\n", __FUNCTION__, __LINE__, #op, pi32, val); \
952	} while (0)
953	MYCHECK(ASMAtomicAddS32(pi32, 1), 10, 11);
954	MYCHECK(ASMAtomicAddS32(pi32, -2), 11, 9);
955	MYCHECK(ASMAtomicAddS32(pi32, -9), 9, 0);
956	MYCHECK(ASMAtomicAddS32(pi32, -0x7fffffff), 0, -0x7fffffff);
957	MYCHECK(ASMAtomicAddS32(pi32, 0), -0x7fffffff, -0x7fffffff);
958	MYCHECK(ASMAtomicAddS32(pi32, 0x7fffffff), -0x7fffffff, 0);
959	MYCHECK(ASMAtomicAddS32(pi32, 0), 0, 0);
960	#undef MYCHECK
961	}
962
963
964	static void tstASMAtomicAddS32(void)
965	{
966	DO_SIMPLE_TEST(ASMAtomicAddS32, int32_t);
967	}
968
969
970	DECLINLINE(void) tstASMAtomicUoIncU32Worker(uint32_t volatile *pu32)
971	{
972	*pu32 = 0;
973
974	CHECKOP(ASMAtomicUoIncU32(pu32), UINT32_C(1), "%#x", uint32_t);
975	CHECKVAL(*pu32, UINT32_C(1), "%#x");
976
977	*pu32 = ~UINT32_C(0);
978	CHECKOP(ASMAtomicUoIncU32(pu32), 0, "%#x", uint32_t);
979	CHECKVAL(*pu32, 0, "%#x");
980
981	*pu32 = UINT32_C(0x7fffffff);
982	CHECKOP(ASMAtomicUoIncU32(pu32), UINT32_C(0x80000000), "%#x", uint32_t);
983	CHECKVAL(*pu32, UINT32_C(0x80000000), "%#x");
984	}
985
986
987	static void tstASMAtomicUoIncU32(void)
988	{
989	DO_SIMPLE_TEST(ASMAtomicUoIncU32, uint32_t);
990	}
991
992
993	DECLINLINE(void) tstASMAtomicUoDecU32Worker(uint32_t volatile *pu32)
994	{
995	*pu32 = 0;
996
997	CHECKOP(ASMAtomicUoDecU32(pu32), ~UINT32_C(0), "%#x", uint32_t);
998	CHECKVAL(*pu32, ~UINT32_C(0), "%#x");
999
1000	*pu32 = ~UINT32_C(0);
1001	CHECKOP(ASMAtomicUoDecU32(pu32), UINT32_C(0xfffffffe), "%#x", uint32_t);
1002	CHECKVAL(*pu32, UINT32_C(0xfffffffe), "%#x");
1003
1004	*pu32 = UINT32_C(0x80000000);
1005	CHECKOP(ASMAtomicUoDecU32(pu32), UINT32_C(0x7fffffff), "%#x", uint32_t);
1006	CHECKVAL(*pu32, UINT32_C(0x7fffffff), "%#x");
1007	}
1008
1009
1010	static void tstASMAtomicUoDecU32(void)
1011	{
1012	DO_SIMPLE_TEST(ASMAtomicUoDecU32, uint32_t);
1013	}
1014
1015
1016	DECLINLINE(void) tstASMAtomicAddS64Worker(int64_t volatile *pi64)
1017	{
1018	int64_t i64Rc;
1019	*pi64 = 10;
1020	#define MYCHECK(op, rc, val) \
1021	do { \
1022	i64Rc = op; \
1023	if (i64Rc != (rc)) \
1024	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %llx expected %llx\n", __FUNCTION__, __LINE__, #op, i64Rc, (int64_t)rc); \
1025	if (*pi64 != (val)) \
1026	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi64=%llx expected %llx\n", __FUNCTION__, __LINE__, #op, pi64, (int64_t)(val)); \
1027	} while (0)
1028	MYCHECK(ASMAtomicAddS64(pi64, 1), 10, 11);
1029	MYCHECK(ASMAtomicAddS64(pi64, -2), 11, 9);
1030	MYCHECK(ASMAtomicAddS64(pi64, -9), 9, 0);
1031	MYCHECK(ASMAtomicAddS64(pi64, -INT64_MAX), 0, -INT64_MAX);
1032	MYCHECK(ASMAtomicAddS64(pi64, 0), -INT64_MAX, -INT64_MAX);
1033	MYCHECK(ASMAtomicAddS64(pi64, -1), -INT64_MAX, INT64_MIN);
1034	MYCHECK(ASMAtomicAddS64(pi64, INT64_MAX), INT64_MIN, -1);
1035	MYCHECK(ASMAtomicAddS64(pi64, 1), -1, 0);
1036	MYCHECK(ASMAtomicAddS64(pi64, 0), 0, 0);
1037	#undef MYCHECK
1038	}
1039
1040
1041	static void tstASMAtomicAddS64(void)
1042	{
1043	DO_SIMPLE_TEST(ASMAtomicAddS64, int64_t);
1044	}
1045
1046
1047	DECLINLINE(void) tstASMAtomicDecIncS32Worker(int32_t volatile *pi32)
1048	{
1049	int32_t i32Rc;
1050	*pi32 = 10;
1051	#define MYCHECK(op, rc) \
1052	do { \
1053	i32Rc = op; \
1054	if (i32Rc != (rc)) \
1055	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %d expected %d\n", __FUNCTION__, __LINE__, #op, i32Rc, rc); \
1056	if (*pi32 != (rc)) \
1057	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi32=%d expected %d\n", __FUNCTION__, __LINE__, #op, pi32, rc); \
1058	} while (0)
1059	MYCHECK(ASMAtomicDecS32(pi32), 9);
1060	MYCHECK(ASMAtomicDecS32(pi32), 8);
1061	MYCHECK(ASMAtomicDecS32(pi32), 7);
1062	MYCHECK(ASMAtomicDecS32(pi32), 6);
1063	MYCHECK(ASMAtomicDecS32(pi32), 5);
1064	MYCHECK(ASMAtomicDecS32(pi32), 4);
1065	MYCHECK(ASMAtomicDecS32(pi32), 3);
1066	MYCHECK(ASMAtomicDecS32(pi32), 2);
1067	MYCHECK(ASMAtomicDecS32(pi32), 1);
1068	MYCHECK(ASMAtomicDecS32(pi32), 0);
1069	MYCHECK(ASMAtomicDecS32(pi32), -1);
1070	MYCHECK(ASMAtomicDecS32(pi32), -2);
1071	MYCHECK(ASMAtomicIncS32(pi32), -1);
1072	MYCHECK(ASMAtomicIncS32(pi32), 0);
1073	MYCHECK(ASMAtomicIncS32(pi32), 1);
1074	MYCHECK(ASMAtomicIncS32(pi32), 2);
1075	MYCHECK(ASMAtomicIncS32(pi32), 3);
1076	MYCHECK(ASMAtomicDecS32(pi32), 2);
1077	MYCHECK(ASMAtomicIncS32(pi32), 3);
1078	MYCHECK(ASMAtomicDecS32(pi32), 2);
1079	MYCHECK(ASMAtomicIncS32(pi32), 3);
1080	#undef MYCHECK
1081	}
1082
1083
1084	static void tstASMAtomicDecIncS32(void)
1085	{
1086	DO_SIMPLE_TEST(ASMAtomicDecIncS32, int32_t);
1087	}
1088
1089
1090	DECLINLINE(void) tstASMAtomicDecIncS64Worker(int64_t volatile *pi64)
1091	{
1092	int64_t i64Rc;
1093	*pi64 = 10;
1094	#define MYCHECK(op, rc) \
1095	do { \
1096	i64Rc = op; \
1097	if (i64Rc != (rc)) \
1098	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %lld expected %lld\n", __FUNCTION__, __LINE__, #op, i64Rc, rc); \
1099	if (*pi64 != (rc)) \
1100	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi64=%lld expected %lld\n", __FUNCTION__, __LINE__, #op, pi64, rc); \
1101	} while (0)
1102	MYCHECK(ASMAtomicDecS64(pi64), 9);
1103	MYCHECK(ASMAtomicDecS64(pi64), 8);
1104	MYCHECK(ASMAtomicDecS64(pi64), 7);
1105	MYCHECK(ASMAtomicDecS64(pi64), 6);
1106	MYCHECK(ASMAtomicDecS64(pi64), 5);
1107	MYCHECK(ASMAtomicDecS64(pi64), 4);
1108	MYCHECK(ASMAtomicDecS64(pi64), 3);
1109	MYCHECK(ASMAtomicDecS64(pi64), 2);
1110	MYCHECK(ASMAtomicDecS64(pi64), 1);
1111	MYCHECK(ASMAtomicDecS64(pi64), 0);
1112	MYCHECK(ASMAtomicDecS64(pi64), -1);
1113	MYCHECK(ASMAtomicDecS64(pi64), -2);
1114	MYCHECK(ASMAtomicIncS64(pi64), -1);
1115	MYCHECK(ASMAtomicIncS64(pi64), 0);
1116	MYCHECK(ASMAtomicIncS64(pi64), 1);
1117	MYCHECK(ASMAtomicIncS64(pi64), 2);
1118	MYCHECK(ASMAtomicIncS64(pi64), 3);
1119	MYCHECK(ASMAtomicDecS64(pi64), 2);
1120	MYCHECK(ASMAtomicIncS64(pi64), 3);
1121	MYCHECK(ASMAtomicDecS64(pi64), 2);
1122	MYCHECK(ASMAtomicIncS64(pi64), 3);
1123	#undef MYCHECK
1124	}
1125
1126
1127	static void tstASMAtomicDecIncS64(void)
1128	{
1129	DO_SIMPLE_TEST(ASMAtomicDecIncS64, int64_t);
1130	}
1131
1132
1133	DECLINLINE(void) tstASMAtomicAndOrU32Worker(uint32_t volatile *pu32)
1134	{
1135	*pu32 = UINT32_C(0xffffffff);
1136
1137	ASMAtomicOrU32(pu32, UINT32_C(0xffffffff));
1138	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
1139
1140	ASMAtomicAndU32(pu32, UINT32_C(0xffffffff));
1141	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
1142
1143	ASMAtomicAndU32(pu32, UINT32_C(0x8f8f8f8f));
1144	CHECKVAL(*pu32, UINT32_C(0x8f8f8f8f), "%x");
1145
1146	ASMAtomicOrU32(pu32, UINT32_C(0x70707070));
1147	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
1148
1149	ASMAtomicAndU32(pu32, UINT32_C(1));
1150	CHECKVAL(*pu32, UINT32_C(1), "%x");
1151
1152	ASMAtomicOrU32(pu32, UINT32_C(0x80000000));
1153	CHECKVAL(*pu32, UINT32_C(0x80000001), "%x");
1154
1155	ASMAtomicAndU32(pu32, UINT32_C(0x80000000));
1156	CHECKVAL(*pu32, UINT32_C(0x80000000), "%x");
1157
1158	ASMAtomicAndU32(pu32, UINT32_C(0));
1159	CHECKVAL(*pu32, UINT32_C(0), "%x");
1160
1161	ASMAtomicOrU32(pu32, UINT32_C(0x42424242));
1162	CHECKVAL(*pu32, UINT32_C(0x42424242), "%x");
1163	}
1164
1165
1166	static void tstASMAtomicAndOrU32(void)
1167	{
1168	DO_SIMPLE_TEST(ASMAtomicAndOrU32, uint32_t);
1169	}
1170
1171
1172	DECLINLINE(void) tstASMAtomicAndOrU64Worker(uint64_t volatile *pu64)
1173	{
1174	*pu64 = UINT64_C(0xffffffff);
1175
1176	ASMAtomicOrU64(pu64, UINT64_C(0xffffffff));
1177	CHECKVAL(*pu64, UINT64_C(0xffffffff), "%x");
1178
1179	ASMAtomicAndU64(pu64, UINT64_C(0xffffffff));
1180	CHECKVAL(*pu64, UINT64_C(0xffffffff), "%x");
1181
1182	ASMAtomicAndU64(pu64, UINT64_C(0x8f8f8f8f));
1183	CHECKVAL(*pu64, UINT64_C(0x8f8f8f8f), "%x");
1184
1185	ASMAtomicOrU64(pu64, UINT64_C(0x70707070));
1186	CHECKVAL(*pu64, UINT64_C(0xffffffff), "%x");
1187
1188	ASMAtomicAndU64(pu64, UINT64_C(1));
1189	CHECKVAL(*pu64, UINT64_C(1), "%x");
1190
1191	ASMAtomicOrU64(pu64, UINT64_C(0x80000000));
1192	CHECKVAL(*pu64, UINT64_C(0x80000001), "%x");
1193
1194	ASMAtomicAndU64(pu64, UINT64_C(0x80000000));
1195	CHECKVAL(*pu64, UINT64_C(0x80000000), "%x");
1196
1197	ASMAtomicAndU64(pu64, UINT64_C(0));
1198	CHECKVAL(*pu64, UINT64_C(0), "%x");
1199
1200	ASMAtomicOrU64(pu64, UINT64_C(0x42424242));
1201	CHECKVAL(*pu64, UINT64_C(0x42424242), "%x");
1202
1203	// Same as above, but now 64-bit wide.
1204	ASMAtomicAndU64(pu64, UINT64_C(0));
1205	CHECKVAL(*pu64, UINT64_C(0), "%x");
1206
1207	ASMAtomicOrU64(pu64, UINT64_C(0xffffffffffffffff));
1208	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%x");
1209
1210	ASMAtomicAndU64(pu64, UINT64_C(0xffffffffffffffff));
1211	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%x");
1212
1213	ASMAtomicAndU64(pu64, UINT64_C(0x8f8f8f8f8f8f8f8f));
1214	CHECKVAL(*pu64, UINT64_C(0x8f8f8f8f8f8f8f8f), "%x");
1215
1216	ASMAtomicOrU64(pu64, UINT64_C(0x7070707070707070));
1217	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%x");
1218
1219	ASMAtomicAndU64(pu64, UINT64_C(1));
1220	CHECKVAL(*pu64, UINT64_C(1), "%x");
1221
1222	ASMAtomicOrU64(pu64, UINT64_C(0x8000000000000000));
1223	CHECKVAL(*pu64, UINT64_C(0x8000000000000001), "%x");
1224
1225	ASMAtomicAndU64(pu64, UINT64_C(0x8000000000000000));
1226	CHECKVAL(*pu64, UINT64_C(0x8000000000000000), "%x");
1227
1228	ASMAtomicAndU64(pu64, UINT64_C(0));
1229	CHECKVAL(*pu64, UINT64_C(0), "%x");
1230
1231	ASMAtomicOrU64(pu64, UINT64_C(0x4242424242424242));
1232	CHECKVAL(*pu64, UINT64_C(0x4242424242424242), "%x");
1233	}
1234
1235
1236	static void tstASMAtomicAndOrU64(void)
1237	{
1238	DO_SIMPLE_TEST(ASMAtomicAndOrU64, uint64_t);
1239	}
1240
1241
1242	DECLINLINE(void) tstASMAtomicUoAndOrU32Worker(uint32_t volatile *pu32)
1243	{
1244	*pu32 = UINT32_C(0xffffffff);
1245
1246	ASMAtomicUoOrU32(pu32, UINT32_C(0xffffffff));
1247	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%#x");
1248
1249	ASMAtomicUoAndU32(pu32, UINT32_C(0xffffffff));
1250	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%#x");
1251
1252	ASMAtomicUoAndU32(pu32, UINT32_C(0x8f8f8f8f));
1253	CHECKVAL(*pu32, UINT32_C(0x8f8f8f8f), "%#x");
1254
1255	ASMAtomicUoOrU32(pu32, UINT32_C(0x70707070));
1256	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%#x");
1257
1258	ASMAtomicUoAndU32(pu32, UINT32_C(1));
1259	CHECKVAL(*pu32, UINT32_C(1), "%#x");
1260
1261	ASMAtomicUoOrU32(pu32, UINT32_C(0x80000000));
1262	CHECKVAL(*pu32, UINT32_C(0x80000001), "%#x");
1263
1264	ASMAtomicUoAndU32(pu32, UINT32_C(0x80000000));
1265	CHECKVAL(*pu32, UINT32_C(0x80000000), "%#x");
1266
1267	ASMAtomicUoAndU32(pu32, UINT32_C(0));
1268	CHECKVAL(*pu32, UINT32_C(0), "%#x");
1269
1270	ASMAtomicUoOrU32(pu32, UINT32_C(0x42424242));
1271	CHECKVAL(*pu32, UINT32_C(0x42424242), "%#x");
1272	}
1273
1274
1275	static void tstASMAtomicUoAndOrU32(void)
1276	{
1277	DO_SIMPLE_TEST(ASMAtomicUoAndOrU32, uint32_t);
1278	}
1279
1280
1281	typedef struct
1282	{
1283	uint8_t ab[PAGE_SIZE];
1284	} TSTPAGE;
1285
1286
1287	DECLINLINE(void) tstASMMemZeroPageWorker(TSTPAGE *pPage)
1288	{
1289	for (unsigned j = 0; j < 16; j++)
1290	{
1291	memset(pPage, 0x11 * j, sizeof(*pPage));
1292	ASMMemZeroPage(pPage);
1293	for (unsigned i = 0; i < sizeof(pPage->ab); i++)
1294	if (pPage->ab[i])
1295	RTTestFailed(g_hTest, "ASMMemZeroPage didn't clear byte at offset %#x!\n", i);
1296	}
1297	}
1298
1299
1300	static void tstASMMemZeroPage(void)
1301	{
1302	DO_SIMPLE_TEST(ASMMemZeroPage, TSTPAGE);
1303	}
1304
1305
1306	void tstASMMemIsZeroPage(RTTEST hTest)
1307	{
1308	RTTestSub(hTest, "ASMMemIsZeroPage");
1309
1310	void *pvPage1 = RTTestGuardedAllocHead(hTest, PAGE_SIZE);
1311	void *pvPage2 = RTTestGuardedAllocTail(hTest, PAGE_SIZE);
1312	RTTESTI_CHECK_RETV(pvPage1 && pvPage2);
1313
1314	memset(pvPage1, 0, PAGE_SIZE);
1315	memset(pvPage2, 0, PAGE_SIZE);
1316	RTTESTI_CHECK(ASMMemIsZeroPage(pvPage1));
1317	RTTESTI_CHECK(ASMMemIsZeroPage(pvPage2));
1318
1319	memset(pvPage1, 0xff, PAGE_SIZE);
1320	memset(pvPage2, 0xff, PAGE_SIZE);
1321	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage1));
1322	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage2));
1323
1324	memset(pvPage1, 0, PAGE_SIZE);
1325	memset(pvPage2, 0, PAGE_SIZE);
1326	for (unsigned off = 0; off < PAGE_SIZE; off++)
1327	{
1328	((uint8_t *)pvPage1)[off] = 1;
1329	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage1));
1330	((uint8_t *)pvPage1)[off] = 0;
1331
1332	((uint8_t *)pvPage2)[off] = 0x80;
1333	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage2));
1334	((uint8_t *)pvPage2)[off] = 0;
1335	}
1336
1337	RTTestSubDone(hTest);
1338	}
1339
1340
1341	void tstASMMemFirstMismatchingU8(RTTEST hTest)
1342	{
1343	RTTestSub(hTest, "ASMMemFirstMismatchingU8");
1344
1345	uint8_t pbPage1 = (uint8_t )RTTestGuardedAllocHead(hTest, PAGE_SIZE);
1346	uint8_t pbPage2 = (uint8_t )RTTestGuardedAllocTail(hTest, PAGE_SIZE);
1347	RTTESTI_CHECK_RETV(pbPage1 && pbPage2);
1348
1349	memset(pbPage1, 0, PAGE_SIZE);
1350	memset(pbPage2, 0, PAGE_SIZE);
1351	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0) == NULL);
1352	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0) == NULL);
1353	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 1) == pbPage1);
1354	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 1) == pbPage2);
1355	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0x87) == pbPage1);
1356	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0x87) == pbPage2);
1357	RTTESTI_CHECK(ASMMemIsZero(pbPage1, PAGE_SIZE));
1358	RTTESTI_CHECK(ASMMemIsZero(pbPage2, PAGE_SIZE));
1359	RTTESTI_CHECK(ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0));
1360	RTTESTI_CHECK(ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0));
1361	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0x34));
1362	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0x88));
1363	unsigned cbSub = 32;
1364	while (cbSub-- > 0)
1365	{
1366	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0) == NULL);
1367	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0) == NULL);
1368	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0) == NULL);
1369	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0) == NULL);
1370
1371	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0x34) == &pbPage1[PAGE_SIZE - cbSub] \|\| !cbSub);
1372	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0x99) == &pbPage2[PAGE_SIZE - cbSub] \|\| !cbSub);
1373	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0x42) == pbPage1 \|\| !cbSub);
1374	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0x88) == pbPage2 \|\| !cbSub);
1375	}
1376
1377	memset(pbPage1, 0xff, PAGE_SIZE);
1378	memset(pbPage2, 0xff, PAGE_SIZE);
1379	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0xff) == NULL);
1380	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0xff) == NULL);
1381	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, PAGE_SIZE, 0xfe) == pbPage1);
1382	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, PAGE_SIZE, 0xfe) == pbPage2);
1383	RTTESTI_CHECK(!ASMMemIsZero(pbPage1, PAGE_SIZE));
1384	RTTESTI_CHECK(!ASMMemIsZero(pbPage2, PAGE_SIZE));
1385	RTTESTI_CHECK(ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0xff));
1386	RTTESTI_CHECK(ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0xff));
1387	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage1, PAGE_SIZE, 0));
1388	RTTESTI_CHECK(!ASMMemIsAllU8(pbPage2, PAGE_SIZE, 0));
1389	cbSub = 32;
1390	while (cbSub-- > 0)
1391	{
1392	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0xff) == NULL);
1393	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0xff) == NULL);
1394	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0xff) == NULL);
1395	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0xff) == NULL);
1396
1397	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage1[PAGE_SIZE - cbSub], cbSub, 0xfe) == &pbPage1[PAGE_SIZE - cbSub] \|\| !cbSub);
1398	RTTESTI_CHECK(ASMMemFirstMismatchingU8(&pbPage2[PAGE_SIZE - cbSub], cbSub, 0xfe) == &pbPage2[PAGE_SIZE - cbSub] \|\| !cbSub);
1399	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage1, cbSub, 0xfe) == pbPage1 \|\| !cbSub);
1400	RTTESTI_CHECK(ASMMemFirstMismatchingU8(pbPage2, cbSub, 0xfe) == pbPage2 \|\| !cbSub);
1401	}
1402
1403
1404	/*
1405	* Various alignments and sizes.
1406	*/
1407	uint8_t const bFiller1 = 0x00;
1408	uint8_t const bFiller2 = 0xf6;
1409	size_t const cbBuf = 128;
1410	uint8_t *pbBuf1 = pbPage1;
1411	uint8_t pbBuf2 = &pbPage2[PAGE_SIZE - cbBuf]; / Put it up against the tail guard */
1412	memset(pbPage1, ~bFiller1, PAGE_SIZE);
1413	memset(pbPage2, ~bFiller2, PAGE_SIZE);
1414	memset(pbBuf1, bFiller1, cbBuf);
1415	memset(pbBuf2, bFiller2, cbBuf);
1416	for (size_t offNonZero = 0; offNonZero < cbBuf; offNonZero++)
1417	{
1418	uint8_t bRand = (uint8_t)RTRandU32();
1419	pbBuf1[offNonZero] = bRand \| 1;
1420	pbBuf2[offNonZero] = (0x80 \| bRand) ^ 0xf6;
1421
1422	for (size_t offStart = 0; offStart < 32; offStart++)
1423	{
1424	size_t const cbMax = cbBuf - offStart;
1425	for (size_t cb = 0; cb < cbMax; cb++)
1426	{
1427	size_t const offEnd = offStart + cb;
1428	uint8_t bSaved1, bSaved2;
1429	if (offEnd < PAGE_SIZE)
1430	{
1431	bSaved1 = pbBuf1[offEnd];
1432	bSaved2 = pbBuf2[offEnd];
1433	pbBuf1[offEnd] = 0xff;
1434	pbBuf2[offEnd] = 0xff;
1435	}
1436
1437	uint8_t pbRet = (uint8_t )ASMMemFirstMismatchingU8(pbBuf1 + offStart, cb, bFiller1);
1438	RTTESTI_CHECK(offNonZero - offStart < cb ? pbRet == &pbBuf1[offNonZero] : pbRet == NULL);
1439
1440	pbRet = (uint8_t *)ASMMemFirstMismatchingU8(pbBuf2 + offStart, cb, bFiller2);
1441	RTTESTI_CHECK(offNonZero - offStart < cb ? pbRet == &pbBuf2[offNonZero] : pbRet == NULL);
1442
1443	if (offEnd < PAGE_SIZE)
1444	{
1445	pbBuf1[offEnd] = bSaved1;
1446	pbBuf2[offEnd] = bSaved2;
1447	}
1448	}
1449	}
1450
1451	pbBuf1[offNonZero] = 0;
1452	pbBuf2[offNonZero] = 0xf6;
1453	}
1454
1455	RTTestSubDone(hTest);
1456	}
1457
1458
1459	void tstASMMemZero32(void)
1460	{
1461	RTTestSub(g_hTest, "ASMMemFill32");
1462
1463	struct
1464	{
1465	uint64_t u64Magic1;
1466	uint8_t abPage[PAGE_SIZE - 32];
1467	uint64_t u64Magic2;
1468	} Buf1, Buf2, Buf3;
1469
1470	Buf1.u64Magic1 = UINT64_C(0xffffffffffffffff);
1471	memset(Buf1.abPage, 0x55, sizeof(Buf1.abPage));
1472	Buf1.u64Magic2 = UINT64_C(0xffffffffffffffff);
1473	Buf2.u64Magic1 = UINT64_C(0xffffffffffffffff);
1474	memset(Buf2.abPage, 0x77, sizeof(Buf2.abPage));
1475	Buf2.u64Magic2 = UINT64_C(0xffffffffffffffff);
1476	Buf3.u64Magic1 = UINT64_C(0xffffffffffffffff);
1477	memset(Buf3.abPage, 0x99, sizeof(Buf3.abPage));
1478	Buf3.u64Magic2 = UINT64_C(0xffffffffffffffff);
1479	ASMMemZero32(Buf1.abPage, sizeof(Buf1.abPage));
1480	ASMMemZero32(Buf2.abPage, sizeof(Buf2.abPage));
1481	ASMMemZero32(Buf3.abPage, sizeof(Buf3.abPage));
1482	if ( Buf1.u64Magic1 != UINT64_C(0xffffffffffffffff)
1483	\|\| Buf1.u64Magic2 != UINT64_C(0xffffffffffffffff)
1484	\|\| Buf2.u64Magic1 != UINT64_C(0xffffffffffffffff)
1485	\|\| Buf2.u64Magic2 != UINT64_C(0xffffffffffffffff)
1486	\|\| Buf3.u64Magic1 != UINT64_C(0xffffffffffffffff)
1487	\|\| Buf3.u64Magic2 != UINT64_C(0xffffffffffffffff))
1488	{
1489	RTTestFailed(g_hTest, "ASMMemZero32 violated one/both magic(s)!\n");
1490	}
1491	for (unsigned i = 0; i < RT_ELEMENTS(Buf1.abPage); i++)
1492	if (Buf1.abPage[i])
1493	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
1494	for (unsigned i = 0; i < RT_ELEMENTS(Buf2.abPage); i++)
1495	if (Buf2.abPage[i])
1496	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
1497	for (unsigned i = 0; i < RT_ELEMENTS(Buf3.abPage); i++)
1498	if (Buf3.abPage[i])
1499	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
1500	}
1501
1502
1503	void tstASMMemFill32(void)
1504	{
1505	RTTestSub(g_hTest, "ASMMemFill32");
1506
1507	struct
1508	{
1509	uint64_t u64Magic1;
1510	uint32_t au32Page[PAGE_SIZE / 4];
1511	uint64_t u64Magic2;
1512	} Buf1;
1513	struct
1514	{
1515	uint64_t u64Magic1;
1516	uint32_t au32Page[(PAGE_SIZE / 4) - 3];
1517	uint64_t u64Magic2;
1518	} Buf2;
1519	struct
1520	{
1521	uint64_t u64Magic1;
1522	uint32_t au32Page[(PAGE_SIZE / 4) - 1];
1523	uint64_t u64Magic2;
1524	} Buf3;
1525
1526	Buf1.u64Magic1 = UINT64_C(0xffffffffffffffff);
1527	memset(Buf1.au32Page, 0x55, sizeof(Buf1.au32Page));
1528	Buf1.u64Magic2 = UINT64_C(0xffffffffffffffff);
1529	Buf2.u64Magic1 = UINT64_C(0xffffffffffffffff);
1530	memset(Buf2.au32Page, 0x77, sizeof(Buf2.au32Page));
1531	Buf2.u64Magic2 = UINT64_C(0xffffffffffffffff);
1532	Buf3.u64Magic1 = UINT64_C(0xffffffffffffffff);
1533	memset(Buf3.au32Page, 0x99, sizeof(Buf3.au32Page));
1534	Buf3.u64Magic2 = UINT64_C(0xffffffffffffffff);
1535	ASMMemFill32(Buf1.au32Page, sizeof(Buf1.au32Page), 0xdeadbeef);
1536	ASMMemFill32(Buf2.au32Page, sizeof(Buf2.au32Page), 0xcafeff01);
1537	ASMMemFill32(Buf3.au32Page, sizeof(Buf3.au32Page), 0xf00dd00f);
1538	if ( Buf1.u64Magic1 != UINT64_C(0xffffffffffffffff)
1539	\|\| Buf1.u64Magic2 != UINT64_C(0xffffffffffffffff)
1540	\|\| Buf2.u64Magic1 != UINT64_C(0xffffffffffffffff)
1541	\|\| Buf2.u64Magic2 != UINT64_C(0xffffffffffffffff)
1542	\|\| Buf3.u64Magic1 != UINT64_C(0xffffffffffffffff)
1543	\|\| Buf3.u64Magic2 != UINT64_C(0xffffffffffffffff))
1544	RTTestFailed(g_hTest, "ASMMemFill32 violated one/both magic(s)!\n");
1545	for (unsigned i = 0; i < RT_ELEMENTS(Buf1.au32Page); i++)
1546	if (Buf1.au32Page[i] != 0xdeadbeef)
1547	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf1.au32Page[i], 0xdeadbeef);
1548	for (unsigned i = 0; i < RT_ELEMENTS(Buf2.au32Page); i++)
1549	if (Buf2.au32Page[i] != 0xcafeff01)
1550	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf2.au32Page[i], 0xcafeff01);
1551	for (unsigned i = 0; i < RT_ELEMENTS(Buf3.au32Page); i++)
1552	if (Buf3.au32Page[i] != 0xf00dd00f)
1553	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf3.au32Page[i], 0xf00dd00f);
1554	}
1555
1556
1557
1558	void tstASMMath(void)
1559	{
1560	RTTestSub(g_hTest, "Math");
1561
1562	uint64_t u64 = ASMMult2xU32RetU64(UINT32_C(0x80000000), UINT32_C(0x10000000));
1563	CHECKVAL(u64, UINT64_C(0x0800000000000000), "%#018RX64");
1564
1565	uint32_t u32 = ASMDivU64ByU32RetU32(UINT64_C(0x0800000000000000), UINT32_C(0x10000000));
1566	CHECKVAL(u32, UINT32_C(0x80000000), "%#010RX32");
1567
1568	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x00000001), UINT32_C(0x00000001), UINT32_C(0x00000001));
1569	CHECKVAL(u32, UINT32_C(0x00000001), "%#018RX32");
1570	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x10000000), UINT32_C(0x80000000), UINT32_C(0x20000000));
1571	CHECKVAL(u32, UINT32_C(0x40000000), "%#018RX32");
1572	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x76543210), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
1573	CHECKVAL(u32, UINT32_C(0x76543210), "%#018RX32");
1574	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0xffffffff), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
1575	CHECKVAL(u32, UINT32_C(0xffffffff), "%#018RX32");
1576	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0xffffffff), UINT32_C(0xfffffff0), UINT32_C(0xffffffff));
1577	CHECKVAL(u32, UINT32_C(0xfffffff0), "%#018RX32");
1578	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x10359583), UINT32_C(0x58734981), UINT32_C(0xf8694045));
1579	CHECKVAL(u32, UINT32_C(0x05c584ce), "%#018RX32");
1580	u32 = ASMMultU32ByU32DivByU32(UINT32_C(0x10359583), UINT32_C(0xf8694045), UINT32_C(0x58734981));
1581	CHECKVAL(u32, UINT32_C(0x2d860795), "%#018RX32");
1582
1583	#if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1584	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x0000000000000001), UINT32_C(0x00000001), UINT32_C(0x00000001));
1585	CHECKVAL(u64, UINT64_C(0x0000000000000001), "%#018RX64");
1586	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x0000000100000000), UINT32_C(0x80000000), UINT32_C(0x00000002));
1587	CHECKVAL(u64, UINT64_C(0x4000000000000000), "%#018RX64");
1588	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xfedcba9876543210), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
1589	CHECKVAL(u64, UINT64_C(0xfedcba9876543210), "%#018RX64");
1590	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xffffffffffffffff), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
1591	CHECKVAL(u64, UINT64_C(0xffffffffffffffff), "%#018RX64");
1592	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xffffffffffffffff), UINT32_C(0xfffffff0), UINT32_C(0xffffffff));
1593	CHECKVAL(u64, UINT64_C(0xfffffff0fffffff0), "%#018RX64");
1594	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x3415934810359583), UINT32_C(0x58734981), UINT32_C(0xf8694045));
1595	CHECKVAL(u64, UINT64_C(0x128b9c3d43184763), "%#018RX64");
1596	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x3415934810359583), UINT32_C(0xf8694045), UINT32_C(0x58734981));
1597	CHECKVAL(u64, UINT64_C(0x924719355cd35a27), "%#018RX64");
1598
1599	# if 0 /* bird: question is whether this should trap or not:
1600	*
1601	* frank: Of course it must trap:
1602	*
1603	* 0xfffffff8 * 0x77d7daf8 = 0x77d7daf441412840
1604	*
1605	* During the following division, the quotient must fit into a 32-bit register.
1606	* Therefore the smallest valid divisor is
1607	*
1608	* (0x77d7daf441412840 >> 32) + 1 = 0x77d7daf5
1609	*
1610	* which is definitely greater than 0x3b9aca00.
1611	*
1612	* bird: No, the C version does not crash. So, the question is whether there's any
1613	* code depending on it not crashing.
1614	*
1615	* Of course the assembly versions of the code crash right now for the reasons you've
1616	* given, but the 32-bit MSC version does not crash.
1617	*
1618	* frank: The C version does not crash but delivers incorrect results for this case.
1619	* The reason is
1620	*
1621	* u.s.Hi = (unsigned long)(u64Hi / u32C);
1622	*
1623	* Here the division is actually 64-bit by 64-bit but the 64-bit result is truncated
1624	* to 32 bit. If using this (optimized and fast) function we should just be sure that
1625	* the operands are in a valid range.
1626	*/
1627	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xfffffff8c65d6731), UINT32_C(0x77d7daf8), UINT32_C(0x3b9aca00));
1628	CHECKVAL(u64, UINT64_C(0x02b8f9a2aa74e3dc), "%#018RX64");
1629	# endif
1630	#endif /* AMD64 \|\| X86 */
1631
1632	u32 = ASMModU64ByU32RetU32(UINT64_C(0x0ffffff8c65d6731), UINT32_C(0x77d7daf8));
1633	CHECKVAL(u32, UINT32_C(0x3B642451), "%#010RX32");
1634
1635	int32_t i32;
1636	i32 = ASMModS64ByS32RetS32(INT64_C(-11), INT32_C(-2));
1637	CHECKVAL(i32, INT32_C(-1), "%010RI32");
1638	i32 = ASMModS64ByS32RetS32(INT64_C(-11), INT32_C(2));
1639	CHECKVAL(i32, INT32_C(-1), "%010RI32");
1640	i32 = ASMModS64ByS32RetS32(INT64_C(11), INT32_C(-2));
1641	CHECKVAL(i32, INT32_C(1), "%010RI32");
1642
1643	i32 = ASMModS64ByS32RetS32(INT64_C(92233720368547758), INT32_C(2147483647));
1644	CHECKVAL(i32, INT32_C(2104533974), "%010RI32");
1645	i32 = ASMModS64ByS32RetS32(INT64_C(-92233720368547758), INT32_C(2147483647));
1646	CHECKVAL(i32, INT32_C(-2104533974), "%010RI32");
1647	}
1648
1649
1650	void tstASMByteSwap(void)
1651	{
1652	RTTestSub(g_hTest, "ASMByteSwap*");
1653
1654	uint64_t u64In = UINT64_C(0x0011223344556677);
1655	uint64_t u64Out = ASMByteSwapU64(u64In);
1656	CHECKVAL(u64In, UINT64_C(0x0011223344556677), "%#018RX64");
1657	CHECKVAL(u64Out, UINT64_C(0x7766554433221100), "%#018RX64");
1658	u64Out = ASMByteSwapU64(u64Out);
1659	CHECKVAL(u64Out, u64In, "%#018RX64");
1660	u64In = UINT64_C(0x0123456789abcdef);
1661	u64Out = ASMByteSwapU64(u64In);
1662	CHECKVAL(u64In, UINT64_C(0x0123456789abcdef), "%#018RX64");
1663	CHECKVAL(u64Out, UINT64_C(0xefcdab8967452301), "%#018RX64");
1664	u64Out = ASMByteSwapU64(u64Out);
1665	CHECKVAL(u64Out, u64In, "%#018RX64");
1666	u64In = 0;
1667	u64Out = ASMByteSwapU64(u64In);
1668	CHECKVAL(u64Out, u64In, "%#018RX64");
1669	u64In = ~(uint64_t)0;
1670	u64Out = ASMByteSwapU64(u64In);
1671	CHECKVAL(u64Out, u64In, "%#018RX64");
1672
1673	uint32_t u32In = UINT32_C(0x00112233);
1674	uint32_t u32Out = ASMByteSwapU32(u32In);
1675	CHECKVAL(u32In, UINT32_C(0x00112233), "%#010RX32");
1676	CHECKVAL(u32Out, UINT32_C(0x33221100), "%#010RX32");
1677	u32Out = ASMByteSwapU32(u32Out);
1678	CHECKVAL(u32Out, u32In, "%#010RX32");
1679	u32In = UINT32_C(0x12345678);
1680	u32Out = ASMByteSwapU32(u32In);
1681	CHECKVAL(u32In, UINT32_C(0x12345678), "%#010RX32");
1682	CHECKVAL(u32Out, UINT32_C(0x78563412), "%#010RX32");
1683	u32Out = ASMByteSwapU32(u32Out);
1684	CHECKVAL(u32Out, u32In, "%#010RX32");
1685	u32In = 0;
1686	u32Out = ASMByteSwapU32(u32In);
1687	CHECKVAL(u32Out, u32In, "%#010RX32");
1688	u32In = ~(uint32_t)0;
1689	u32Out = ASMByteSwapU32(u32In);
1690	CHECKVAL(u32Out, u32In, "%#010RX32");
1691
1692	uint16_t u16In = UINT16_C(0x0011);
1693	uint16_t u16Out = ASMByteSwapU16(u16In);
1694	CHECKVAL(u16In, UINT16_C(0x0011), "%#06RX16");
1695	CHECKVAL(u16Out, UINT16_C(0x1100), "%#06RX16");
1696	u16Out = ASMByteSwapU16(u16Out);
1697	CHECKVAL(u16Out, u16In, "%#06RX16");
1698	u16In = UINT16_C(0x1234);
1699	u16Out = ASMByteSwapU16(u16In);
1700	CHECKVAL(u16In, UINT16_C(0x1234), "%#06RX16");
1701	CHECKVAL(u16Out, UINT16_C(0x3412), "%#06RX16");
1702	u16Out = ASMByteSwapU16(u16Out);
1703	CHECKVAL(u16Out, u16In, "%#06RX16");
1704	u16In = 0;
1705	u16Out = ASMByteSwapU16(u16In);
1706	CHECKVAL(u16Out, u16In, "%#06RX16");
1707	u16In = ~(uint16_t)0;
1708	u16Out = ASMByteSwapU16(u16In);
1709	CHECKVAL(u16Out, u16In, "%#06RX16");
1710	}
1711
1712
1713	void tstASMBench(void)
1714	{
1715	/*
1716	* Make this static. We don't want to have this located on the stack.
1717	*/
1718	static uint8_t volatile s_u8;
1719	static int8_t volatile s_i8;
1720	static uint16_t volatile s_u16;
1721	static int16_t volatile s_i16;
1722	static uint32_t volatile s_u32;
1723	static int32_t volatile s_i32;
1724	static uint64_t volatile s_u64;
1725	static int64_t volatile s_i64;
1726	register unsigned i;
1727	const unsigned cRounds = _2M; /* Must be multiple of 8 */
1728	register uint64_t u64Elapsed;
1729
1730	RTTestSub(g_hTest, "Benchmarking");
1731
1732	#if 0 && !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
1733	# define BENCH(op, str) \
1734	do { \
1735	RTThreadYield(); \
1736	u64Elapsed = ASMReadTSC(); \
1737	for (i = cRounds; i > 0; i--) \
1738	op; \
1739	u64Elapsed = ASMReadTSC() - u64Elapsed; \
1740	RTTestValue(g_hTest, str, u64Elapsed / cRounds, RTTESTUNIT_TICKS_PER_CALL); \
1741	} while (0)
1742	#else
1743	# define BENCH(op, str) \
1744	do { \
1745	RTThreadYield(); \
1746	u64Elapsed = RTTimeNanoTS(); \
1747	for (i = cRounds / 8; i > 0; i--) \
1748	{ \
1749	op; \
1750	op; \
1751	op; \
1752	op; \
1753	op; \
1754	op; \
1755	op; \
1756	op; \
1757	} \
1758	u64Elapsed = RTTimeNanoTS() - u64Elapsed; \
1759	RTTestValue(g_hTest, str, u64Elapsed / cRounds, RTTESTUNIT_NS_PER_CALL); \
1760	} while (0)
1761	#endif
1762	#if (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)) && !defined(GCC44_32BIT_PIC)
1763	# define BENCH_TSC(op, str) \
1764	do { \
1765	RTThreadYield(); \
1766	u64Elapsed = ASMReadTSC(); \
1767	for (i = cRounds / 8; i > 0; i--) \
1768	{ \
1769	op; \
1770	op; \
1771	op; \
1772	op; \
1773	op; \
1774	op; \
1775	op; \
1776	op; \
1777	} \
1778	u64Elapsed = ASMReadTSC() - u64Elapsed; \
1779	RTTestValue(g_hTest, str, u64Elapsed / cRounds, /RTTESTUNIT_TICKS_PER_CALL/ RTTESTUNIT_NONE); \
1780	} while (0)
1781	#else
1782	# define BENCH_TSC(op, str) BENCH(op, str)
1783	#endif
1784
1785	BENCH(s_u32 = 0, "s_u32 = 0");
1786	BENCH(ASMAtomicUoReadU8(&s_u8), "ASMAtomicUoReadU8");
1787	BENCH(ASMAtomicUoReadS8(&s_i8), "ASMAtomicUoReadS8");
1788	BENCH(ASMAtomicUoReadU16(&s_u16), "ASMAtomicUoReadU16");
1789	BENCH(ASMAtomicUoReadS16(&s_i16), "ASMAtomicUoReadS16");
1790	BENCH(ASMAtomicUoReadU32(&s_u32), "ASMAtomicUoReadU32");
1791	BENCH(ASMAtomicUoReadS32(&s_i32), "ASMAtomicUoReadS32");
1792	BENCH(ASMAtomicUoReadU64(&s_u64), "ASMAtomicUoReadU64");
1793	BENCH(ASMAtomicUoReadS64(&s_i64), "ASMAtomicUoReadS64");
1794	BENCH(ASMAtomicReadU8(&s_u8), "ASMAtomicReadU8");
1795	BENCH(ASMAtomicReadS8(&s_i8), "ASMAtomicReadS8");
1796	BENCH(ASMAtomicReadU16(&s_u16), "ASMAtomicReadU16");
1797	BENCH(ASMAtomicReadS16(&s_i16), "ASMAtomicReadS16");
1798	BENCH(ASMAtomicReadU32(&s_u32), "ASMAtomicReadU32");
1799	BENCH(ASMAtomicReadS32(&s_i32), "ASMAtomicReadS32");
1800	BENCH(ASMAtomicReadU64(&s_u64), "ASMAtomicReadU64");
1801	BENCH(ASMAtomicReadS64(&s_i64), "ASMAtomicReadS64");
1802	BENCH(ASMAtomicUoWriteU8(&s_u8, 0), "ASMAtomicUoWriteU8");
1803	BENCH(ASMAtomicUoWriteS8(&s_i8, 0), "ASMAtomicUoWriteS8");
1804	BENCH(ASMAtomicUoWriteU16(&s_u16, 0), "ASMAtomicUoWriteU16");
1805	BENCH(ASMAtomicUoWriteS16(&s_i16, 0), "ASMAtomicUoWriteS16");
1806	BENCH(ASMAtomicUoWriteU32(&s_u32, 0), "ASMAtomicUoWriteU32");
1807	BENCH(ASMAtomicUoWriteS32(&s_i32, 0), "ASMAtomicUoWriteS32");
1808	BENCH(ASMAtomicUoWriteU64(&s_u64, 0), "ASMAtomicUoWriteU64");
1809	BENCH(ASMAtomicUoWriteS64(&s_i64, 0), "ASMAtomicUoWriteS64");
1810	BENCH(ASMAtomicWriteU8(&s_u8, 0), "ASMAtomicWriteU8");
1811	BENCH(ASMAtomicWriteS8(&s_i8, 0), "ASMAtomicWriteS8");
1812	BENCH(ASMAtomicWriteU16(&s_u16, 0), "ASMAtomicWriteU16");
1813	BENCH(ASMAtomicWriteS16(&s_i16, 0), "ASMAtomicWriteS16");
1814	BENCH(ASMAtomicWriteU32(&s_u32, 0), "ASMAtomicWriteU32");
1815	BENCH(ASMAtomicWriteS32(&s_i32, 0), "ASMAtomicWriteS32");
1816	BENCH(ASMAtomicWriteU64(&s_u64, 0), "ASMAtomicWriteU64");
1817	BENCH(ASMAtomicWriteS64(&s_i64, 0), "ASMAtomicWriteS64");
1818	BENCH(ASMAtomicXchgU8(&s_u8, 0), "ASMAtomicXchgU8");
1819	BENCH(ASMAtomicXchgS8(&s_i8, 0), "ASMAtomicXchgS8");
1820	BENCH(ASMAtomicXchgU16(&s_u16, 0), "ASMAtomicXchgU16");
1821	BENCH(ASMAtomicXchgS16(&s_i16, 0), "ASMAtomicXchgS16");
1822	BENCH(ASMAtomicXchgU32(&s_u32, 0), "ASMAtomicXchgU32");
1823	BENCH(ASMAtomicXchgS32(&s_i32, 0), "ASMAtomicXchgS32");
1824	BENCH(ASMAtomicXchgU64(&s_u64, 0), "ASMAtomicXchgU64");
1825	BENCH(ASMAtomicXchgS64(&s_i64, 0), "ASMAtomicXchgS64");
1826	BENCH(ASMAtomicCmpXchgU32(&s_u32, 0, 0), "ASMAtomicCmpXchgU32");
1827	BENCH(ASMAtomicCmpXchgS32(&s_i32, 0, 0), "ASMAtomicCmpXchgS32");
1828	BENCH(ASMAtomicCmpXchgU64(&s_u64, 0, 0), "ASMAtomicCmpXchgU64");
1829	BENCH(ASMAtomicCmpXchgS64(&s_i64, 0, 0), "ASMAtomicCmpXchgS64");
1830	BENCH(ASMAtomicCmpXchgU32(&s_u32, 0, 1), "ASMAtomicCmpXchgU32/neg");
1831	BENCH(ASMAtomicCmpXchgS32(&s_i32, 0, 1), "ASMAtomicCmpXchgS32/neg");
1832	BENCH(ASMAtomicCmpXchgU64(&s_u64, 0, 1), "ASMAtomicCmpXchgU64/neg");
1833	BENCH(ASMAtomicCmpXchgS64(&s_i64, 0, 1), "ASMAtomicCmpXchgS64/neg");
1834	BENCH(ASMAtomicIncU32(&s_u32), "ASMAtomicIncU32");
1835	BENCH(ASMAtomicIncS32(&s_i32), "ASMAtomicIncS32");
1836	BENCH(ASMAtomicDecU32(&s_u32), "ASMAtomicDecU32");
1837	BENCH(ASMAtomicDecS32(&s_i32), "ASMAtomicDecS32");
1838	BENCH(ASMAtomicAddU32(&s_u32, 5), "ASMAtomicAddU32");
1839	BENCH(ASMAtomicAddS32(&s_i32, 5), "ASMAtomicAddS32");
1840	BENCH(ASMAtomicUoIncU32(&s_u32), "ASMAtomicUoIncU32");
1841	BENCH(ASMAtomicUoDecU32(&s_u32), "ASMAtomicUoDecU32");
1842	BENCH(ASMAtomicUoAndU32(&s_u32, 0xffffffff), "ASMAtomicUoAndU32");
1843	BENCH(ASMAtomicUoOrU32(&s_u32, 0xffffffff), "ASMAtomicUoOrU32");
1844	BENCH_TSC(ASMSerializeInstructionCpuId(), "ASMSerializeInstructionCpuId");
1845	BENCH_TSC(ASMSerializeInstructionIRet(), "ASMSerializeInstructionIRet");
1846
1847	/* The Darwin gcc does not like this ... */
1848	#if !defined(RT_OS_DARWIN) && !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
1849	BENCH(s_u8 = ASMGetApicId(), "ASMGetApicId");
1850	#endif
1851	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
1852	uint32_t uAux;
1853	if ( ASMHasCpuId()
1854	&& ASMIsValidExtRange(ASMCpuId_EAX(0x80000000))
1855	&& (ASMCpuId_EDX(0x80000001) & X86_CPUID_EXT_FEATURE_EDX_RDTSCP) )
1856	{
1857	BENCH_TSC(ASMSerializeInstructionRdTscp(), "ASMSerializeInstructionRdTscp");
1858	BENCH(s_u64 = ASMReadTscWithAux(&uAux), "ASMReadTscWithAux");
1859	}
1860	BENCH(s_u64 = ASMReadTSC(), "ASMReadTSC");
1861	union
1862	{
1863	uint64_t u64[2];
1864	RTIDTR Unaligned;
1865	struct
1866	{
1867	uint16_t abPadding[3];
1868	RTIDTR Aligned;
1869	} s;
1870	} uBuf;
1871	Assert(((uintptr_t)&uBuf.Unaligned.pIdt & (sizeof(uintptr_t) - 1)) != 0);
1872	BENCH(ASMGetIDTR(&uBuf.Unaligned), "ASMGetIDTR/unaligned");
1873	Assert(((uintptr_t)&uBuf.s.Aligned.pIdt & (sizeof(uintptr_t) - 1)) == 0);
1874	BENCH(ASMGetIDTR(&uBuf.s.Aligned), "ASMGetIDTR/aligned");
1875	#endif
1876
1877	#undef BENCH
1878	}
1879
1880
1881	int main(int argc, char *argv[])
1882	{
1883	int rc = RTTestInitAndCreate("tstRTInlineAsm", &g_hTest);
1884	if (rc)
1885	return rc;
1886	RTTestBanner(g_hTest);
1887
1888	/*
1889	* Execute the tests.
1890	*/
1891	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
1892	tstASMCpuId();
1893	//bruteForceCpuId();
1894	#endif
1895	#if 1
1896	tstASMAtomicXchgU8();
1897	tstASMAtomicXchgU16();
1898	tstASMAtomicXchgU32();
1899	tstASMAtomicXchgU64();
1900	tstASMAtomicXchgPtr();
1901	tstASMAtomicCmpXchgU8();
1902	tstASMAtomicCmpXchgU32();
1903	tstASMAtomicCmpXchgU64();
1904	tstASMAtomicCmpXchgExU32();
1905	tstASMAtomicCmpXchgExU64();
1906	tstASMAtomicReadU64();
1907	tstASMAtomicUoReadU64();
1908
1909	tstASMAtomicAddS32();
1910	tstASMAtomicAddS64();
1911	tstASMAtomicDecIncS32();
1912	tstASMAtomicDecIncS64();
1913	tstASMAtomicAndOrU32();
1914	tstASMAtomicAndOrU64();
1915
1916	tstASMAtomicUoIncU32();
1917	tstASMAtomicUoDecU32();
1918	tstASMAtomicUoAndOrU32();
1919
1920	tstASMMemZeroPage();
1921	tstASMMemIsZeroPage(g_hTest);
1922	tstASMMemFirstMismatchingU8(g_hTest);
1923	tstASMMemZero32();
1924	tstASMMemFill32();
1925
1926	tstASMMath();
1927
1928	tstASMByteSwap();
1929
1930	tstASMBench();
1931	#endif
1932
1933	/*
1934	* Show the result.
1935	*/
1936	return RTTestSummaryAndDestroy(g_hTest);
1937	}
1938

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

source: vbox/trunk/src/VBox/Runtime/testcase/tstRTInlineAsm.cpp@ 61311

Download in other formats: