tstRTInlineAsm.cpp@ 36363

Last change on this file since 36363 was 33207, checked in by vboxsync, 14 years ago
tstRTInlineAsm: another fix
Property svn:eol-style set to `native` Property svn:keywords set to `Id`
File size: 58.3 KB

Line
1	/* $Id: tstRTInlineAsm.cpp 33207 2010-10-18 15:02:47Z vboxsync $ */
2	/** @file
3	* IPRT Testcase - inline assembly.
4	*/
5
6	/*
7	* Copyright (C) 2006-2010 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	* Header Files *
29	*******************************************************************************/
30	#include <iprt/asm.h>
31	#include <iprt/asm-math.h>
32
33	/* See http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44018. Only gcc version 4.4
34	* is affected. No harm for the VBox code: If the cpuid code compiles, it works
35	* fine. */
36	#if defined(__GNUC__) && defined(RT_ARCH_X86) && defined(__PIC__)
37	# if __GNUC__ == 4 && __GNUC_MINOR__ == 4
38	# define GCC44_32BIT_PIC
39	# endif
40	#endif
41
42	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
43	# include <iprt/asm-amd64-x86.h>
44	#else
45	# include <iprt/time.h>
46	#endif
47	#include <iprt/stream.h>
48	#include <iprt/string.h>
49	#include <iprt/param.h>
50	#include <iprt/thread.h>
51	#include <iprt/test.h>
52	#include <iprt/time.h>
53
54
55
56	/*******************************************************************************
57	* Defined Constants And Macros *
58	*******************************************************************************/
59	#define CHECKVAL(val, expect, fmt) \
60	do \
61	{ \
62	if ((val) != (expect)) \
63	{ \
64	RTTestFailed(g_hTest, "%s, %d: " #val ": expected " fmt " got " fmt "\n", __FUNCTION__, __LINE__, (expect), (val)); \
65	} \
66	} while (0)
67
68	#define CHECKOP(op, expect, fmt, type) \
69	do \
70	{ \
71	type val = op; \
72	if (val != (type)(expect)) \
73	{ \
74	RTTestFailed(g_hTest, "%s, %d: " #op ": expected " fmt " got " fmt "\n", __FUNCTION__, __LINE__, (type)(expect), val); \
75	} \
76	} while (0)
77
78	/**
79	* Calls a worker function with different worker variable storage types.
80	*/
81	#define DO_SIMPLE_TEST(name, type) \
82	do \
83	{ \
84	RTTestISub(#name); \
85	type StackVar; \
86	tst ## name ## Worker(&StackVar); \
87	\
88	type pVar = (type )RTTestGuardedAllocHead(g_hTest, sizeof(type)); \
89	RTTEST_CHECK_BREAK(g_hTest, pVar); \
90	tst ## name ## Worker(pVar); \
91	RTTestGuardedFree(g_hTest, pVar); \
92	\
93	pVar = (type *)RTTestGuardedAllocTail(g_hTest, sizeof(type)); \
94	RTTEST_CHECK_BREAK(g_hTest, pVar); \
95	tst ## name ## Worker(pVar); \
96	RTTestGuardedFree(g_hTest, pVar); \
97	} while (0)
98
99
100	/*******************************************************************************
101	* Global Variables *
102	*******************************************************************************/
103	/** The test instance. */
104	static RTTEST g_hTest;
105
106
107
108	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
109
110	const char *getCacheAss(unsigned u)
111	{
112	if (u == 0)
113	return "res0 ";
114	if (u == 1)
115	return "direct";
116	if (u >= 256)
117	return "???";
118
119	char *pszRet;
120	RTStrAPrintf(&pszRet, "%d way", u); /* intentional leak! */
121	return pszRet;
122	}
123
124
125	const char *getL2CacheAss(unsigned u)
126	{
127	switch (u)
128	{
129	case 0: return "off ";
130	case 1: return "direct";
131	case 2: return "2 way ";
132	case 3: return "res3 ";
133	case 4: return "4 way ";
134	case 5: return "res5 ";
135	case 6: return "8 way ";
136	case 7: return "res7 ";
137	case 8: return "16 way";
138	case 9: return "res9 ";
139	case 10: return "res10 ";
140	case 11: return "res11 ";
141	case 12: return "res12 ";
142	case 13: return "res13 ";
143	case 14: return "res14 ";
144	case 15: return "fully ";
145	default:
146	return "????";
147	}
148	}
149
150
151	/**
152	* Test and dump all possible info from the CPUID instruction.
153	*
154	* @remark Bits shared with the libc cpuid.c program. This all written by me, so no worries.
155	* @todo transform the dumping into a generic runtime function. We'll need it for logging!
156	*/
157	void tstASMCpuId(void)
158	{
159	RTTestISub("ASMCpuId");
160
161	unsigned iBit;
162	struct
163	{
164	uint32_t uEBX, uEAX, uEDX, uECX;
165	} s;
166	if (!ASMHasCpuId())
167	{
168	RTTestIPrintf(RTTESTLVL_ALWAYS, "warning! CPU doesn't support CPUID\n");
169	return;
170	}
171
172	/*
173	* Try the 0 function and use that for checking the ASMCpuId_* variants.
174	*/
175	ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
176
177	uint32_t u32;
178
179	u32 = ASMCpuId_EAX(0);
180	CHECKVAL(u32, s.uEAX, "%x");
181	u32 = ASMCpuId_EBX(0);
182	CHECKVAL(u32, s.uEBX, "%x");
183	u32 = ASMCpuId_ECX(0);
184	CHECKVAL(u32, s.uECX, "%x");
185	u32 = ASMCpuId_EDX(0);
186	CHECKVAL(u32, s.uEDX, "%x");
187
188	uint32_t uECX2 = s.uECX - 1;
189	uint32_t uEDX2 = s.uEDX - 1;
190	ASMCpuId_ECX_EDX(0, &uECX2, &uEDX2);
191	CHECKVAL(uECX2, s.uECX, "%x");
192	CHECKVAL(uEDX2, s.uEDX, "%x");
193
194	/*
195	* Done testing, dump the information.
196	*/
197	RTTestIPrintf(RTTESTLVL_ALWAYS, "CPUID Dump\n");
198	ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
199	const uint32_t cFunctions = s.uEAX;
200
201	/* raw dump */
202	RTTestIPrintf(RTTESTLVL_ALWAYS,
203	"\n"
204	" RAW Standard CPUIDs\n"
205	"Function eax ebx ecx edx\n");
206	for (unsigned iStd = 0; iStd <= cFunctions + 3; iStd++)
207	{
208	ASMCpuId_Idx_ECX(iStd, 0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
209	RTTestIPrintf(RTTESTLVL_ALWAYS, "%08x %08x %08x %08x %08x%s\n",
210	iStd, s.uEAX, s.uEBX, s.uECX, s.uEDX, iStd <= cFunctions ? "" : "*");
211
212	if (iStd == 0x04 \|\| iStd == 0x0b \|\| iStd == 0x0d \|\| iStd > cFunctions)
213	continue; /* Leaf 04 and leaf 0d output depend on the initial value of ECX
214	* The same seems to apply to invalid standard functions */
215
216	u32 = ASMCpuId_EAX(iStd);
217	CHECKVAL(u32, s.uEAX, "%x");
218	u32 = ASMCpuId_EBX(iStd);
219	CHECKVAL(u32, s.uEBX, "%x");
220	u32 = ASMCpuId_ECX(iStd);
221	CHECKVAL(u32, s.uECX, "%x");
222	u32 = ASMCpuId_EDX(iStd);
223	CHECKVAL(u32, s.uEDX, "%x");
224
225	uECX2 = s.uECX - 1;
226	uEDX2 = s.uEDX - 1;
227	ASMCpuId_ECX_EDX(iStd, &uECX2, &uEDX2);
228	CHECKVAL(uECX2, s.uECX, "%x");
229	CHECKVAL(uEDX2, s.uEDX, "%x");
230	}
231
232	/*
233	* Understandable output
234	*/
235	ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
236	RTTestIPrintf(RTTESTLVL_ALWAYS,
237	"Name: %.04s%.04s%.04s\n"
238	"Support: 0-%u\n",
239	&s.uEBX, &s.uEDX, &s.uECX, s.uEAX);
240	bool const fIntel = ASMIsIntelCpuEx(s.uEBX, s.uECX, s.uEDX);
241
242	/*
243	* Get Features.
244	*/
245	if (cFunctions >= 1)
246	{
247	static const char * const s_apszTypes[4] = { "primary", "overdrive", "MP", "reserved" };
248	ASMCpuId(1, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
249	RTTestIPrintf(RTTESTLVL_ALWAYS,
250	"Family: %#x \tExtended: %#x \tEffective: %#x\n"
251	"Model: %#x \tExtended: %#x \tEffective: %#x\n"
252	"Stepping: %d\n"
253	"Type: %d (%s)\n"
254	"APIC ID: %#04x\n"
255	"Logical CPUs: %d\n"
256	"CLFLUSH Size: %d\n"
257	"Brand ID: %#04x\n",
258	(s.uEAX >> 8) & 0xf, (s.uEAX >> 20) & 0x7f, ASMGetCpuFamily(s.uEAX),
259	(s.uEAX >> 4) & 0xf, (s.uEAX >> 16) & 0x0f, ASMGetCpuModel(s.uEAX, fIntel),
260	ASMGetCpuStepping(s.uEAX),
261	(s.uEAX >> 12) & 0x3, s_apszTypes[(s.uEAX >> 12) & 0x3],
262	(s.uEBX >> 24) & 0xff,
263	(s.uEBX >> 16) & 0xff,
264	(s.uEBX >> 8) & 0xff,
265	(s.uEBX >> 0) & 0xff);
266
267	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features EDX: ");
268	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FPU");
269	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VME");
270	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DE");
271	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE");
272	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TSC");
273	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MSR");
274	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAE");
275	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCE");
276	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CX8");
277	if (s.uEDX & RT_BIT(9)) RTTestIPrintf(RTTESTLVL_ALWAYS, " APIC");
278	if (s.uEDX & RT_BIT(10)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 10");
279	if (s.uEDX & RT_BIT(11)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SEP");
280	if (s.uEDX & RT_BIT(12)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MTRR");
281	if (s.uEDX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PGE");
282	if (s.uEDX & RT_BIT(14)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCA");
283	if (s.uEDX & RT_BIT(15)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMOV");
284	if (s.uEDX & RT_BIT(16)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAT");
285	if (s.uEDX & RT_BIT(17)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE36");
286	if (s.uEDX & RT_BIT(18)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSN");
287	if (s.uEDX & RT_BIT(19)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CLFSH");
288	if (s.uEDX & RT_BIT(20)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 20");
289	if (s.uEDX & RT_BIT(21)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DS");
290	if (s.uEDX & RT_BIT(22)) RTTestIPrintf(RTTESTLVL_ALWAYS, " ACPI");
291	if (s.uEDX & RT_BIT(23)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MMX");
292	if (s.uEDX & RT_BIT(24)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FXSR");
293	if (s.uEDX & RT_BIT(25)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE");
294	if (s.uEDX & RT_BIT(26)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE2");
295	if (s.uEDX & RT_BIT(27)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SS");
296	if (s.uEDX & RT_BIT(28)) RTTestIPrintf(RTTESTLVL_ALWAYS, " HTT");
297	if (s.uEDX & RT_BIT(29)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 29");
298	if (s.uEDX & RT_BIT(30)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 30");
299	if (s.uEDX & RT_BIT(31)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 31");
300	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
301
302	/** @todo check intel docs. */
303	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features ECX: ");
304	if (s.uECX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SSE3");
305	for (iBit = 1; iBit < 13; iBit++)
306	if (s.uECX & RT_BIT(iBit))
307	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
308	if (s.uECX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CX16");
309	for (iBit = 14; iBit < 32; iBit++)
310	if (s.uECX & RT_BIT(iBit))
311	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
312	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
313	}
314
315	/*
316	* Extended.
317	* Implemented after AMD specs.
318	*/
319	/** @todo check out the intel specs. */
320	ASMCpuId(0x80000000, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
321	if (!s.uEAX && !s.uEBX && !s.uECX && !s.uEDX)
322	{
323	RTTestIPrintf(RTTESTLVL_ALWAYS, "No extended CPUID info? Check the manual on how to detect this...\n");
324	return;
325	}
326	const uint32_t cExtFunctions = s.uEAX \| 0x80000000;
327
328	/* raw dump */
329	RTTestIPrintf(RTTESTLVL_ALWAYS,
330	"\n"
331	" RAW Extended CPUIDs\n"
332	"Function eax ebx ecx edx\n");
333	for (unsigned iExt = 0x80000000; iExt <= cExtFunctions + 3; iExt++)
334	{
335	ASMCpuId(iExt, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
336	RTTestIPrintf(RTTESTLVL_ALWAYS, "%08x %08x %08x %08x %08x%s\n",
337	iExt, s.uEAX, s.uEBX, s.uECX, s.uEDX, iExt <= cExtFunctions ? "" : "*");
338
339	if (iExt > cExtFunctions)
340	continue; /* Invalid extended functions seems change the value if ECX changes */
341
342	u32 = ASMCpuId_EAX(iExt);
343	CHECKVAL(u32, s.uEAX, "%x");
344	u32 = ASMCpuId_EBX(iExt);
345	CHECKVAL(u32, s.uEBX, "%x");
346	u32 = ASMCpuId_ECX(iExt);
347	CHECKVAL(u32, s.uECX, "%x");
348	u32 = ASMCpuId_EDX(iExt);
349	CHECKVAL(u32, s.uEDX, "%x");
350
351	uECX2 = s.uECX - 1;
352	uEDX2 = s.uEDX - 1;
353	ASMCpuId_ECX_EDX(iExt, &uECX2, &uEDX2);
354	CHECKVAL(uECX2, s.uECX, "%x");
355	CHECKVAL(uEDX2, s.uEDX, "%x");
356	}
357
358	/*
359	* Understandable output
360	*/
361	ASMCpuId(0x80000000, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
362	RTTestIPrintf(RTTESTLVL_ALWAYS,
363	"Ext Name: %.4s%.4s%.4s\n"
364	"Ext Supports: 0x80000000-%#010x\n",
365	&s.uEBX, &s.uEDX, &s.uECX, s.uEAX);
366
367	if (cExtFunctions >= 0x80000001)
368	{
369	ASMCpuId(0x80000001, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
370	RTTestIPrintf(RTTESTLVL_ALWAYS,
371	"Family: %#x \tExtended: %#x \tEffective: %#x\n"
372	"Model: %#x \tExtended: %#x \tEffective: %#x\n"
373	"Stepping: %d\n"
374	"Brand ID: %#05x\n",
375	(s.uEAX >> 8) & 0xf, (s.uEAX >> 20) & 0x7f, ASMGetCpuFamily(s.uEAX),
376	(s.uEAX >> 4) & 0xf, (s.uEAX >> 16) & 0x0f, ASMGetCpuModel(s.uEAX, fIntel),
377	ASMGetCpuStepping(s.uEAX),
378	s.uEBX & 0xfff);
379
380	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features EDX: ");
381	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FPU");
382	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VME");
383	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " DE");
384	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE");
385	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TSC");
386	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MSR");
387	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAE");
388	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCE");
389	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMPXCHG8B");
390	if (s.uEDX & RT_BIT(9)) RTTestIPrintf(RTTESTLVL_ALWAYS, " APIC");
391	if (s.uEDX & RT_BIT(10)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 10");
392	if (s.uEDX & RT_BIT(11)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SysCallSysRet");
393	if (s.uEDX & RT_BIT(12)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MTRR");
394	if (s.uEDX & RT_BIT(13)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PGE");
395	if (s.uEDX & RT_BIT(14)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MCA");
396	if (s.uEDX & RT_BIT(15)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CMOV");
397	if (s.uEDX & RT_BIT(16)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PAT");
398	if (s.uEDX & RT_BIT(17)) RTTestIPrintf(RTTESTLVL_ALWAYS, " PSE36");
399	if (s.uEDX & RT_BIT(18)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 18");
400	if (s.uEDX & RT_BIT(19)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 19");
401	if (s.uEDX & RT_BIT(20)) RTTestIPrintf(RTTESTLVL_ALWAYS, " NX");
402	if (s.uEDX & RT_BIT(21)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 21");
403	if (s.uEDX & RT_BIT(22)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MmxExt");
404	if (s.uEDX & RT_BIT(23)) RTTestIPrintf(RTTESTLVL_ALWAYS, " MMX");
405	if (s.uEDX & RT_BIT(24)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FXSR");
406	if (s.uEDX & RT_BIT(25)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FastFXSR");
407	if (s.uEDX & RT_BIT(26)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 26");
408	if (s.uEDX & RT_BIT(27)) RTTestIPrintf(RTTESTLVL_ALWAYS, " RDTSCP");
409	if (s.uEDX & RT_BIT(28)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 28");
410	if (s.uEDX & RT_BIT(29)) RTTestIPrintf(RTTESTLVL_ALWAYS, " LongMode");
411	if (s.uEDX & RT_BIT(30)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3DNowExt");
412	if (s.uEDX & RT_BIT(31)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3DNow");
413	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
414
415	RTTestIPrintf(RTTESTLVL_ALWAYS, "Features ECX: ");
416	if (s.uECX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " LahfSahf");
417	if (s.uECX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " CmpLegacy");
418	if (s.uECX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " SVM");
419	if (s.uECX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 3");
420	if (s.uECX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " AltMovCr8");
421	for (iBit = 5; iBit < 32; iBit++)
422	if (s.uECX & RT_BIT(iBit))
423	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
424	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
425	}
426
427	char szString[443+1] = {0};
428	if (cExtFunctions >= 0x80000002)
429	ASMCpuId(0x80000002, &szString[0 + 0], &szString[0 + 4], &szString[0 + 8], &szString[0 + 12]);
430	if (cExtFunctions >= 0x80000003)
431	ASMCpuId(0x80000003, &szString[16 + 0], &szString[16 + 4], &szString[16 + 8], &szString[16 + 12]);
432	if (cExtFunctions >= 0x80000004)
433	ASMCpuId(0x80000004, &szString[32 + 0], &szString[32 + 4], &szString[32 + 8], &szString[32 + 12]);
434	if (cExtFunctions >= 0x80000002)
435	RTTestIPrintf(RTTESTLVL_ALWAYS, "Full Name: %s\n", szString);
436
437	if (cExtFunctions >= 0x80000005)
438	{
439	ASMCpuId(0x80000005, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
440	RTTestIPrintf(RTTESTLVL_ALWAYS,
441	"TLB 2/4M Instr/Uni: %s %3d entries\n"
442	"TLB 2/4M Data: %s %3d entries\n",
443	getCacheAss((s.uEAX >> 8) & 0xff), (s.uEAX >> 0) & 0xff,
444	getCacheAss((s.uEAX >> 24) & 0xff), (s.uEAX >> 16) & 0xff);
445	RTTestIPrintf(RTTESTLVL_ALWAYS,
446	"TLB 4K Instr/Uni: %s %3d entries\n"
447	"TLB 4K Data: %s %3d entries\n",
448	getCacheAss((s.uEBX >> 8) & 0xff), (s.uEBX >> 0) & 0xff,
449	getCacheAss((s.uEBX >> 24) & 0xff), (s.uEBX >> 16) & 0xff);
450	RTTestIPrintf(RTTESTLVL_ALWAYS,
451	"L1 Instr Cache Line Size: %d bytes\n"
452	"L1 Instr Cache Lines Per Tag: %d\n"
453	"L1 Instr Cache Associativity: %s\n"
454	"L1 Instr Cache Size: %d KB\n",
455	(s.uEDX >> 0) & 0xff,
456	(s.uEDX >> 8) & 0xff,
457	getCacheAss((s.uEDX >> 16) & 0xff),
458	(s.uEDX >> 24) & 0xff);
459	RTTestIPrintf(RTTESTLVL_ALWAYS,
460	"L1 Data Cache Line Size: %d bytes\n"
461	"L1 Data Cache Lines Per Tag: %d\n"
462	"L1 Data Cache Associativity: %s\n"
463	"L1 Data Cache Size: %d KB\n",
464	(s.uECX >> 0) & 0xff,
465	(s.uECX >> 8) & 0xff,
466	getCacheAss((s.uECX >> 16) & 0xff),
467	(s.uECX >> 24) & 0xff);
468	}
469
470	if (cExtFunctions >= 0x80000006)
471	{
472	ASMCpuId(0x80000006, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
473	RTTestIPrintf(RTTESTLVL_ALWAYS,
474	"L2 TLB 2/4M Instr/Uni: %s %4d entries\n"
475	"L2 TLB 2/4M Data: %s %4d entries\n",
476	getL2CacheAss((s.uEAX >> 12) & 0xf), (s.uEAX >> 0) & 0xfff,
477	getL2CacheAss((s.uEAX >> 28) & 0xf), (s.uEAX >> 16) & 0xfff);
478	RTTestIPrintf(RTTESTLVL_ALWAYS,
479	"L2 TLB 4K Instr/Uni: %s %4d entries\n"
480	"L2 TLB 4K Data: %s %4d entries\n",
481	getL2CacheAss((s.uEBX >> 12) & 0xf), (s.uEBX >> 0) & 0xfff,
482	getL2CacheAss((s.uEBX >> 28) & 0xf), (s.uEBX >> 16) & 0xfff);
483	RTTestIPrintf(RTTESTLVL_ALWAYS,
484	"L2 Cache Line Size: %d bytes\n"
485	"L2 Cache Lines Per Tag: %d\n"
486	"L2 Cache Associativity: %s\n"
487	"L2 Cache Size: %d KB\n",
488	(s.uEDX >> 0) & 0xff,
489	(s.uEDX >> 8) & 0xf,
490	getL2CacheAss((s.uEDX >> 12) & 0xf),
491	(s.uEDX >> 16) & 0xffff);
492	}
493
494	if (cExtFunctions >= 0x80000007)
495	{
496	ASMCpuId(0x80000007, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
497	RTTestIPrintf(RTTESTLVL_ALWAYS, "APM Features: ");
498	if (s.uEDX & RT_BIT(0)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TS");
499	if (s.uEDX & RT_BIT(1)) RTTestIPrintf(RTTESTLVL_ALWAYS, " FID");
500	if (s.uEDX & RT_BIT(2)) RTTestIPrintf(RTTESTLVL_ALWAYS, " VID");
501	if (s.uEDX & RT_BIT(3)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TTP");
502	if (s.uEDX & RT_BIT(4)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TM");
503	if (s.uEDX & RT_BIT(5)) RTTestIPrintf(RTTESTLVL_ALWAYS, " STC");
504	if (s.uEDX & RT_BIT(6)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 6");
505	if (s.uEDX & RT_BIT(7)) RTTestIPrintf(RTTESTLVL_ALWAYS, " 7");
506	if (s.uEDX & RT_BIT(8)) RTTestIPrintf(RTTESTLVL_ALWAYS, " TscInvariant");
507	for (iBit = 9; iBit < 32; iBit++)
508	if (s.uEDX & RT_BIT(iBit))
509	RTTestIPrintf(RTTESTLVL_ALWAYS, " %d", iBit);
510	RTTestIPrintf(RTTESTLVL_ALWAYS, "\n");
511	}
512
513	if (cExtFunctions >= 0x80000008)
514	{
515	ASMCpuId(0x80000008, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
516	RTTestIPrintf(RTTESTLVL_ALWAYS,
517	"Physical Address Width: %d bits\n"
518	"Virtual Address Width: %d bits\n"
519	"Guest Physical Address Width: %d bits\n",
520	(s.uEAX >> 0) & 0xff,
521	(s.uEAX >> 8) & 0xff,
522	(s.uEAX >> 16) & 0xff);
523	RTTestIPrintf(RTTESTLVL_ALWAYS,
524	"Physical Core Count: %d\n",
525	((s.uECX >> 0) & 0xff) + 1);
526	if ((s.uECX >> 12) & 0xf)
527	RTTestIPrintf(RTTESTLVL_ALWAYS, "ApicIdCoreIdSize: %d bits\n", (s.uECX >> 12) & 0xf);
528	}
529
530	if (cExtFunctions >= 0x8000000a)
531	{
532	ASMCpuId(0x8000000a, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
533	RTTestIPrintf(RTTESTLVL_ALWAYS,
534	"SVM Revision: %d (%#x)\n"
535	"Number of Address Space IDs: %d (%#x)\n",
536	s.uEAX & 0xff, s.uEAX & 0xff,
537	s.uEBX, s.uEBX);
538	}
539	}
540
541	#endif /* AMD64 \|\| X86 */
542
543	DECLINLINE(void) tstASMAtomicXchgU8Worker(uint8_t volatile *pu8)
544	{
545	*pu8 = 0;
546	CHECKOP(ASMAtomicXchgU8(pu8, 1), 0, "%#x", uint8_t);
547	CHECKVAL(*pu8, 1, "%#x");
548
549	CHECKOP(ASMAtomicXchgU8(pu8, 0), 1, "%#x", uint8_t);
550	CHECKVAL(*pu8, 0, "%#x");
551
552	CHECKOP(ASMAtomicXchgU8(pu8, 0xff), 0, "%#x", uint8_t);
553	CHECKVAL(*pu8, 0xff, "%#x");
554
555	CHECKOP(ASMAtomicXchgU8(pu8, 0x87), 0xffff, "%#x", uint8_t);
556	CHECKVAL(*pu8, 0x87, "%#x");
557	}
558
559
560	static void tstASMAtomicXchgU8(void)
561	{
562	DO_SIMPLE_TEST(ASMAtomicXchgU8, uint8_t);
563	}
564
565
566	DECLINLINE(void) tstASMAtomicXchgU16Worker(uint16_t volatile *pu16)
567	{
568	*pu16 = 0;
569
570	CHECKOP(ASMAtomicXchgU16(pu16, 1), 0, "%#x", uint16_t);
571	CHECKVAL(*pu16, 1, "%#x");
572
573	CHECKOP(ASMAtomicXchgU16(pu16, 0), 1, "%#x", uint16_t);
574	CHECKVAL(*pu16, 0, "%#x");
575
576	CHECKOP(ASMAtomicXchgU16(pu16, 0xffff), 0, "%#x", uint16_t);
577	CHECKVAL(*pu16, 0xffff, "%#x");
578
579	CHECKOP(ASMAtomicXchgU16(pu16, 0x8765), 0xffff, "%#x", uint16_t);
580	CHECKVAL(*pu16, 0x8765, "%#x");
581	}
582
583
584	static void tstASMAtomicXchgU16(void)
585	{
586	DO_SIMPLE_TEST(ASMAtomicXchgU16, uint16_t);
587	}
588
589
590	DECLINLINE(void) tstASMAtomicXchgU32Worker(uint32_t volatile *pu32)
591	{
592	*pu32 = 0;
593
594	CHECKOP(ASMAtomicXchgU32(pu32, 1), 0, "%#x", uint32_t);
595	CHECKVAL(*pu32, 1, "%#x");
596
597	CHECKOP(ASMAtomicXchgU32(pu32, 0), 1, "%#x", uint32_t);
598	CHECKVAL(*pu32, 0, "%#x");
599
600	CHECKOP(ASMAtomicXchgU32(pu32, ~UINT32_C(0)), 0, "%#x", uint32_t);
601	CHECKVAL(*pu32, ~UINT32_C(0), "%#x");
602
603	CHECKOP(ASMAtomicXchgU32(pu32, 0x87654321), ~UINT32_C(0), "%#x", uint32_t);
604	CHECKVAL(*pu32, 0x87654321, "%#x");
605	}
606
607
608	static void tstASMAtomicXchgU32(void)
609	{
610	DO_SIMPLE_TEST(ASMAtomicXchgU32, uint32_t);
611	}
612
613
614	DECLINLINE(void) tstASMAtomicXchgU64Worker(uint64_t volatile *pu64)
615	{
616	*pu64 = 0;
617
618	CHECKOP(ASMAtomicXchgU64(pu64, 1), UINT64_C(0), "%#llx", uint64_t);
619	CHECKVAL(*pu64, UINT64_C(1), "%#llx");
620
621	CHECKOP(ASMAtomicXchgU64(pu64, 0), UINT64_C(1), "%#llx", uint64_t);
622	CHECKVAL(*pu64, UINT64_C(0), "%#llx");
623
624	CHECKOP(ASMAtomicXchgU64(pu64, ~UINT64_C(0)), UINT64_C(0), "%#llx", uint64_t);
625	CHECKVAL(*pu64, ~UINT64_C(0), "%#llx");
626
627	CHECKOP(ASMAtomicXchgU64(pu64, UINT64_C(0xfedcba0987654321)), ~UINT64_C(0), "%#llx", uint64_t);
628	CHECKVAL(*pu64, UINT64_C(0xfedcba0987654321), "%#llx");
629	}
630
631
632	static void tstASMAtomicXchgU64(void)
633	{
634	DO_SIMPLE_TEST(ASMAtomicXchgU64, uint64_t);
635	}
636
637
638	DECLINLINE(void) tstASMAtomicXchgPtrWorker(void * volatile *ppv)
639	{
640	*ppv = NULL;
641
642	CHECKOP(ASMAtomicXchgPtr(ppv, (void )(~(uintptr_t)0)), NULL, "%p", void );
643	CHECKVAL(ppv, (void )(~(uintptr_t)0), "%p");
644
645	CHECKOP(ASMAtomicXchgPtr(ppv, (void )0x87654321), (void )(~(uintptr_t)0), "%p", void *);
646	CHECKVAL(ppv, (void )0x87654321, "%p");
647
648	CHECKOP(ASMAtomicXchgPtr(ppv, NULL), (void )0x87654321, "%p", void );
649	CHECKVAL(*ppv, NULL, "%p");
650	}
651
652
653	static void tstASMAtomicXchgPtr(void)
654	{
655	DO_SIMPLE_TEST(ASMAtomicXchgPtr, void *);
656	}
657
658
659	DECLINLINE(void) tstASMAtomicCmpXchgU8Worker(uint8_t volatile *pu8)
660	{
661	*pu8 = 0xff;
662
663	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0, 0), false, "%d", bool);
664	CHECKVAL(*pu8, 0xff, "%x");
665
666	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0, 0xff), true, "%d", bool);
667	CHECKVAL(*pu8, 0, "%x");
668
669	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0x79, 0xff), false, "%d", bool);
670	CHECKVAL(*pu8, 0, "%x");
671
672	CHECKOP(ASMAtomicCmpXchgU8(pu8, 0x97, 0), true, "%d", bool);
673	CHECKVAL(*pu8, 0x97, "%x");
674	}
675
676
677	static void tstASMAtomicCmpXchgU8(void)
678	{
679	DO_SIMPLE_TEST(ASMAtomicCmpXchgU8, uint8_t);
680	}
681
682
683	DECLINLINE(void) tstASMAtomicCmpXchgU32Worker(uint32_t volatile *pu32)
684	{
685	*pu32 = UINT32_C(0xffffffff);
686
687	CHECKOP(ASMAtomicCmpXchgU32(pu32, 0, 0), false, "%d", bool);
688	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
689
690	CHECKOP(ASMAtomicCmpXchgU32(pu32, 0, UINT32_C(0xffffffff)), true, "%d", bool);
691	CHECKVAL(*pu32, 0, "%x");
692
693	CHECKOP(ASMAtomicCmpXchgU32(pu32, UINT32_C(0x8008efd), UINT32_C(0xffffffff)), false, "%d", bool);
694	CHECKVAL(*pu32, 0, "%x");
695
696	CHECKOP(ASMAtomicCmpXchgU32(pu32, UINT32_C(0x8008efd), 0), true, "%d", bool);
697	CHECKVAL(*pu32, UINT32_C(0x8008efd), "%x");
698	}
699
700
701	static void tstASMAtomicCmpXchgU32(void)
702	{
703	DO_SIMPLE_TEST(ASMAtomicCmpXchgU32, uint32_t);
704	}
705
706
707
708	DECLINLINE(void) tstASMAtomicCmpXchgU64Worker(uint64_t volatile *pu64)
709	{
710	*pu64 = UINT64_C(0xffffffffffffff);
711
712	CHECKOP(ASMAtomicCmpXchgU64(pu64, 0, 0), false, "%d", bool);
713	CHECKVAL(*pu64, UINT64_C(0xffffffffffffff), "%#llx");
714
715	CHECKOP(ASMAtomicCmpXchgU64(pu64, 0, UINT64_C(0xffffffffffffff)), true, "%d", bool);
716	CHECKVAL(*pu64, 0, "%x");
717
718	CHECKOP(ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), UINT64_C(0xffffffff)), false, "%d", bool);
719	CHECKVAL(*pu64, 0, "%x");
720
721	CHECKOP(ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), UINT64_C(0xffffffff00000000)), false, "%d", bool);
722	CHECKVAL(*pu64, 0, "%x");
723
724	CHECKOP(ASMAtomicCmpXchgU64(pu64, UINT64_C(0x80040008008efd), 0), true, "%d", bool);
725	CHECKVAL(*pu64, UINT64_C(0x80040008008efd), "%#llx");
726	}
727
728
729	static void tstASMAtomicCmpXchgU64(void)
730	{
731	DO_SIMPLE_TEST(ASMAtomicCmpXchgU64, uint64_t);
732	}
733
734
735	DECLINLINE(void) tstASMAtomicCmpXchgExU32Worker(uint32_t volatile *pu32)
736	{
737	*pu32 = UINT32_C(0xffffffff);
738	uint32_t u32Old = UINT32_C(0x80005111);
739
740	CHECKOP(ASMAtomicCmpXchgExU32(pu32, 0, 0, &u32Old), false, "%d", bool);
741	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
742	CHECKVAL(u32Old, UINT32_C(0xffffffff), "%x");
743
744	CHECKOP(ASMAtomicCmpXchgExU32(pu32, 0, UINT32_C(0xffffffff), &u32Old), true, "%d", bool);
745	CHECKVAL(*pu32, 0, "%x");
746	CHECKVAL(u32Old, UINT32_C(0xffffffff), "%x");
747
748	CHECKOP(ASMAtomicCmpXchgExU32(pu32, UINT32_C(0x8008efd), UINT32_C(0xffffffff), &u32Old), false, "%d", bool);
749	CHECKVAL(*pu32, 0, "%x");
750	CHECKVAL(u32Old, 0, "%x");
751
752	CHECKOP(ASMAtomicCmpXchgExU32(pu32, UINT32_C(0x8008efd), 0, &u32Old), true, "%d", bool);
753	CHECKVAL(*pu32, UINT32_C(0x8008efd), "%x");
754	CHECKVAL(u32Old, 0, "%x");
755
756	CHECKOP(ASMAtomicCmpXchgExU32(pu32, 0, UINT32_C(0x8008efd), &u32Old), true, "%d", bool);
757	CHECKVAL(*pu32, 0, "%x");
758	CHECKVAL(u32Old, UINT32_C(0x8008efd), "%x");
759	}
760
761
762	static void tstASMAtomicCmpXchgExU32(void)
763	{
764	DO_SIMPLE_TEST(ASMAtomicCmpXchgExU32, uint32_t);
765	}
766
767
768	DECLINLINE(void) tstASMAtomicCmpXchgExU64Worker(uint64_t volatile *pu64)
769	{
770	*pu64 = UINT64_C(0xffffffffffffffff);
771	uint64_t u64Old = UINT64_C(0x8000000051111111);
772
773	CHECKOP(ASMAtomicCmpXchgExU64(pu64, 0, 0, &u64Old), false, "%d", bool);
774	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%llx");
775	CHECKVAL(u64Old, UINT64_C(0xffffffffffffffff), "%llx");
776
777	CHECKOP(ASMAtomicCmpXchgExU64(pu64, 0, UINT64_C(0xffffffffffffffff), &u64Old), true, "%d", bool);
778	CHECKVAL(*pu64, UINT64_C(0), "%llx");
779	CHECKVAL(u64Old, UINT64_C(0xffffffffffffffff), "%llx");
780
781	CHECKOP(ASMAtomicCmpXchgExU64(pu64, UINT64_C(0x80040008008efd), 0xffffffff, &u64Old), false, "%d", bool);
782	CHECKVAL(*pu64, UINT64_C(0), "%llx");
783	CHECKVAL(u64Old, UINT64_C(0), "%llx");
784
785	CHECKOP(ASMAtomicCmpXchgExU64(pu64, UINT64_C(0x80040008008efd), UINT64_C(0xffffffff00000000), &u64Old), false, "%d", bool);
786	CHECKVAL(*pu64, UINT64_C(0), "%llx");
787	CHECKVAL(u64Old, UINT64_C(0), "%llx");
788
789	CHECKOP(ASMAtomicCmpXchgExU64(pu64, UINT64_C(0x80040008008efd), 0, &u64Old), true, "%d", bool);
790	CHECKVAL(*pu64, UINT64_C(0x80040008008efd), "%llx");
791	CHECKVAL(u64Old, UINT64_C(0), "%llx");
792
793	CHECKOP(ASMAtomicCmpXchgExU64(pu64, 0, UINT64_C(0x80040008008efd), &u64Old), true, "%d", bool);
794	CHECKVAL(*pu64, UINT64_C(0), "%llx");
795	CHECKVAL(u64Old, UINT64_C(0x80040008008efd), "%llx");
796	}
797
798
799	static void tstASMAtomicCmpXchgExU64(void)
800	{
801	DO_SIMPLE_TEST(ASMAtomicCmpXchgExU64, uint64_t);
802	}
803
804
805	DECLINLINE(void) tstASMAtomicReadU64Worker(uint64_t volatile *pu64)
806	{
807	*pu64 = 0;
808
809	CHECKOP(ASMAtomicReadU64(pu64), UINT64_C(0), "%#llx", uint64_t);
810	CHECKVAL(*pu64, UINT64_C(0), "%#llx");
811
812	*pu64 = ~UINT64_C(0);
813	CHECKOP(ASMAtomicReadU64(pu64), ~UINT64_C(0), "%#llx", uint64_t);
814	CHECKVAL(*pu64, ~UINT64_C(0), "%#llx");
815
816	*pu64 = UINT64_C(0xfedcba0987654321);
817	CHECKOP(ASMAtomicReadU64(pu64), UINT64_C(0xfedcba0987654321), "%#llx", uint64_t);
818	CHECKVAL(*pu64, UINT64_C(0xfedcba0987654321), "%#llx");
819	}
820
821
822	static void tstASMAtomicReadU64(void)
823	{
824	DO_SIMPLE_TEST(ASMAtomicReadU64, uint64_t);
825	}
826
827
828	DECLINLINE(void) tstASMAtomicUoReadU64Worker(uint64_t volatile *pu64)
829	{
830	*pu64 = 0;
831
832	CHECKOP(ASMAtomicUoReadU64(pu64), UINT64_C(0), "%#llx", uint64_t);
833	CHECKVAL(*pu64, UINT64_C(0), "%#llx");
834
835	*pu64 = ~UINT64_C(0);
836	CHECKOP(ASMAtomicUoReadU64(pu64), ~UINT64_C(0), "%#llx", uint64_t);
837	CHECKVAL(*pu64, ~UINT64_C(0), "%#llx");
838
839	*pu64 = UINT64_C(0xfedcba0987654321);
840	CHECKOP(ASMAtomicUoReadU64(pu64), UINT64_C(0xfedcba0987654321), "%#llx", uint64_t);
841	CHECKVAL(*pu64, UINT64_C(0xfedcba0987654321), "%#llx");
842	}
843
844
845	static void tstASMAtomicUoReadU64(void)
846	{
847	DO_SIMPLE_TEST(ASMAtomicUoReadU64, uint64_t);
848	}
849
850
851	DECLINLINE(void) tstASMAtomicAddS32Worker(int32_t *pi32)
852	{
853	int32_t i32Rc;
854	*pi32 = 10;
855	#define MYCHECK(op, rc, val) \
856	do { \
857	i32Rc = op; \
858	if (i32Rc != (rc)) \
859	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %d expected %d\n", __FUNCTION__, __LINE__, #op, i32Rc, rc); \
860	if (*pi32 != (val)) \
861	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi32=%d expected %d\n", __FUNCTION__, __LINE__, #op, pi32, val); \
862	} while (0)
863	MYCHECK(ASMAtomicAddS32(pi32, 1), 10, 11);
864	MYCHECK(ASMAtomicAddS32(pi32, -2), 11, 9);
865	MYCHECK(ASMAtomicAddS32(pi32, -9), 9, 0);
866	MYCHECK(ASMAtomicAddS32(pi32, -0x7fffffff), 0, -0x7fffffff);
867	MYCHECK(ASMAtomicAddS32(pi32, 0), -0x7fffffff, -0x7fffffff);
868	MYCHECK(ASMAtomicAddS32(pi32, 0x7fffffff), -0x7fffffff, 0);
869	MYCHECK(ASMAtomicAddS32(pi32, 0), 0, 0);
870	#undef MYCHECK
871	}
872
873	static void tstASMAtomicAddS32(void)
874	{
875	DO_SIMPLE_TEST(ASMAtomicAddS32, int32_t);
876	}
877
878
879	DECLINLINE(void) tstASMAtomicAddS64Worker(int64_t volatile *pi64)
880	{
881	int64_t i64Rc;
882	*pi64 = 10;
883	#define MYCHECK(op, rc, val) \
884	do { \
885	i64Rc = op; \
886	if (i64Rc != (rc)) \
887	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %llx expected %llx\n", __FUNCTION__, __LINE__, #op, i64Rc, (int64_t)rc); \
888	if (*pi64 != (val)) \
889	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi64=%llx expected %llx\n", __FUNCTION__, __LINE__, #op, pi64, (int64_t)(val)); \
890	} while (0)
891	MYCHECK(ASMAtomicAddS64(pi64, 1), 10, 11);
892	MYCHECK(ASMAtomicAddS64(pi64, -2), 11, 9);
893	MYCHECK(ASMAtomicAddS64(pi64, -9), 9, 0);
894	MYCHECK(ASMAtomicAddS64(pi64, -INT64_MAX), 0, -INT64_MAX);
895	MYCHECK(ASMAtomicAddS64(pi64, 0), -INT64_MAX, -INT64_MAX);
896	MYCHECK(ASMAtomicAddS64(pi64, -1), -INT64_MAX, INT64_MIN);
897	MYCHECK(ASMAtomicAddS64(pi64, INT64_MAX), INT64_MIN, -1);
898	MYCHECK(ASMAtomicAddS64(pi64, 1), -1, 0);
899	MYCHECK(ASMAtomicAddS64(pi64, 0), 0, 0);
900	#undef MYCHECK
901	}
902
903
904	static void tstASMAtomicAddS64(void)
905	{
906	DO_SIMPLE_TEST(ASMAtomicAddS64, int64_t);
907	}
908
909
910	DECLINLINE(void) tstASMAtomicDecIncS32Worker(int32_t volatile *pi32)
911	{
912	int32_t i32Rc;
913	*pi32 = 10;
914	#define MYCHECK(op, rc) \
915	do { \
916	i32Rc = op; \
917	if (i32Rc != (rc)) \
918	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %d expected %d\n", __FUNCTION__, __LINE__, #op, i32Rc, rc); \
919	if (*pi32 != (rc)) \
920	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi32=%d expected %d\n", __FUNCTION__, __LINE__, #op, pi32, rc); \
921	} while (0)
922	MYCHECK(ASMAtomicDecS32(pi32), 9);
923	MYCHECK(ASMAtomicDecS32(pi32), 8);
924	MYCHECK(ASMAtomicDecS32(pi32), 7);
925	MYCHECK(ASMAtomicDecS32(pi32), 6);
926	MYCHECK(ASMAtomicDecS32(pi32), 5);
927	MYCHECK(ASMAtomicDecS32(pi32), 4);
928	MYCHECK(ASMAtomicDecS32(pi32), 3);
929	MYCHECK(ASMAtomicDecS32(pi32), 2);
930	MYCHECK(ASMAtomicDecS32(pi32), 1);
931	MYCHECK(ASMAtomicDecS32(pi32), 0);
932	MYCHECK(ASMAtomicDecS32(pi32), -1);
933	MYCHECK(ASMAtomicDecS32(pi32), -2);
934	MYCHECK(ASMAtomicIncS32(pi32), -1);
935	MYCHECK(ASMAtomicIncS32(pi32), 0);
936	MYCHECK(ASMAtomicIncS32(pi32), 1);
937	MYCHECK(ASMAtomicIncS32(pi32), 2);
938	MYCHECK(ASMAtomicIncS32(pi32), 3);
939	MYCHECK(ASMAtomicDecS32(pi32), 2);
940	MYCHECK(ASMAtomicIncS32(pi32), 3);
941	MYCHECK(ASMAtomicDecS32(pi32), 2);
942	MYCHECK(ASMAtomicIncS32(pi32), 3);
943	#undef MYCHECK
944	}
945
946
947	static void tstASMAtomicDecIncS32(void)
948	{
949	DO_SIMPLE_TEST(ASMAtomicDecIncS32, int32_t);
950	}
951
952
953	DECLINLINE(void) tstASMAtomicDecIncS64Worker(int64_t volatile *pi64)
954	{
955	int64_t i64Rc;
956	*pi64 = 10;
957	#define MYCHECK(op, rc) \
958	do { \
959	i64Rc = op; \
960	if (i64Rc != (rc)) \
961	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s -> %lld expected %lld\n", __FUNCTION__, __LINE__, #op, i64Rc, rc); \
962	if (*pi64 != (rc)) \
963	RTTestFailed(g_hTest, "%s, %d: FAILURE: %s => pi64=%lld expected %lld\n", __FUNCTION__, __LINE__, #op, pi64, rc); \
964	} while (0)
965	MYCHECK(ASMAtomicDecS64(pi64), 9);
966	MYCHECK(ASMAtomicDecS64(pi64), 8);
967	MYCHECK(ASMAtomicDecS64(pi64), 7);
968	MYCHECK(ASMAtomicDecS64(pi64), 6);
969	MYCHECK(ASMAtomicDecS64(pi64), 5);
970	MYCHECK(ASMAtomicDecS64(pi64), 4);
971	MYCHECK(ASMAtomicDecS64(pi64), 3);
972	MYCHECK(ASMAtomicDecS64(pi64), 2);
973	MYCHECK(ASMAtomicDecS64(pi64), 1);
974	MYCHECK(ASMAtomicDecS64(pi64), 0);
975	MYCHECK(ASMAtomicDecS64(pi64), -1);
976	MYCHECK(ASMAtomicDecS64(pi64), -2);
977	MYCHECK(ASMAtomicIncS64(pi64), -1);
978	MYCHECK(ASMAtomicIncS64(pi64), 0);
979	MYCHECK(ASMAtomicIncS64(pi64), 1);
980	MYCHECK(ASMAtomicIncS64(pi64), 2);
981	MYCHECK(ASMAtomicIncS64(pi64), 3);
982	MYCHECK(ASMAtomicDecS64(pi64), 2);
983	MYCHECK(ASMAtomicIncS64(pi64), 3);
984	MYCHECK(ASMAtomicDecS64(pi64), 2);
985	MYCHECK(ASMAtomicIncS64(pi64), 3);
986	#undef MYCHECK
987	}
988
989
990	static void tstASMAtomicDecIncS64(void)
991	{
992	DO_SIMPLE_TEST(ASMAtomicDecIncS64, int64_t);
993	}
994
995
996	DECLINLINE(void) tstASMAtomicAndOrU32Worker(uint32_t volatile *pu32)
997	{
998	*pu32 = UINT32_C(0xffffffff);
999
1000	ASMAtomicOrU32(pu32, UINT32_C(0xffffffff));
1001	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
1002
1003	ASMAtomicAndU32(pu32, UINT32_C(0xffffffff));
1004	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
1005
1006	ASMAtomicAndU32(pu32, UINT32_C(0x8f8f8f8f));
1007	CHECKVAL(*pu32, UINT32_C(0x8f8f8f8f), "%x");
1008
1009	ASMAtomicOrU32(pu32, UINT32_C(0x70707070));
1010	CHECKVAL(*pu32, UINT32_C(0xffffffff), "%x");
1011
1012	ASMAtomicAndU32(pu32, UINT32_C(1));
1013	CHECKVAL(*pu32, UINT32_C(1), "%x");
1014
1015	ASMAtomicOrU32(pu32, UINT32_C(0x80000000));
1016	CHECKVAL(*pu32, UINT32_C(0x80000001), "%x");
1017
1018	ASMAtomicAndU32(pu32, UINT32_C(0x80000000));
1019	CHECKVAL(*pu32, UINT32_C(0x80000000), "%x");
1020
1021	ASMAtomicAndU32(pu32, UINT32_C(0));
1022	CHECKVAL(*pu32, UINT32_C(0), "%x");
1023
1024	ASMAtomicOrU32(pu32, UINT32_C(0x42424242));
1025	CHECKVAL(*pu32, UINT32_C(0x42424242), "%x");
1026	}
1027
1028
1029	static void tstASMAtomicAndOrU32(void)
1030	{
1031	DO_SIMPLE_TEST(ASMAtomicAndOrU32, uint32_t);
1032	}
1033
1034
1035	DECLINLINE(void) tstASMAtomicAndOrU64Worker(uint64_t volatile *pu64)
1036	{
1037	*pu64 = UINT64_C(0xffffffff);
1038
1039	ASMAtomicOrU64(pu64, UINT64_C(0xffffffff));
1040	CHECKVAL(*pu64, UINT64_C(0xffffffff), "%x");
1041
1042	ASMAtomicAndU64(pu64, UINT64_C(0xffffffff));
1043	CHECKVAL(*pu64, UINT64_C(0xffffffff), "%x");
1044
1045	ASMAtomicAndU64(pu64, UINT64_C(0x8f8f8f8f));
1046	CHECKVAL(*pu64, UINT64_C(0x8f8f8f8f), "%x");
1047
1048	ASMAtomicOrU64(pu64, UINT64_C(0x70707070));
1049	CHECKVAL(*pu64, UINT64_C(0xffffffff), "%x");
1050
1051	ASMAtomicAndU64(pu64, UINT64_C(1));
1052	CHECKVAL(*pu64, UINT64_C(1), "%x");
1053
1054	ASMAtomicOrU64(pu64, UINT64_C(0x80000000));
1055	CHECKVAL(*pu64, UINT64_C(0x80000001), "%x");
1056
1057	ASMAtomicAndU64(pu64, UINT64_C(0x80000000));
1058	CHECKVAL(*pu64, UINT64_C(0x80000000), "%x");
1059
1060	ASMAtomicAndU64(pu64, UINT64_C(0));
1061	CHECKVAL(*pu64, UINT64_C(0), "%x");
1062
1063	ASMAtomicOrU64(pu64, UINT64_C(0x42424242));
1064	CHECKVAL(*pu64, UINT64_C(0x42424242), "%x");
1065
1066	// Same as above, but now 64-bit wide.
1067	ASMAtomicAndU64(pu64, UINT64_C(0));
1068	CHECKVAL(*pu64, UINT64_C(0), "%x");
1069
1070	ASMAtomicOrU64(pu64, UINT64_C(0xffffffffffffffff));
1071	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%x");
1072
1073	ASMAtomicAndU64(pu64, UINT64_C(0xffffffffffffffff));
1074	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%x");
1075
1076	ASMAtomicAndU64(pu64, UINT64_C(0x8f8f8f8f8f8f8f8f));
1077	CHECKVAL(*pu64, UINT64_C(0x8f8f8f8f8f8f8f8f), "%x");
1078
1079	ASMAtomicOrU64(pu64, UINT64_C(0x7070707070707070));
1080	CHECKVAL(*pu64, UINT64_C(0xffffffffffffffff), "%x");
1081
1082	ASMAtomicAndU64(pu64, UINT64_C(1));
1083	CHECKVAL(*pu64, UINT64_C(1), "%x");
1084
1085	ASMAtomicOrU64(pu64, UINT64_C(0x8000000000000000));
1086	CHECKVAL(*pu64, UINT64_C(0x8000000000000001), "%x");
1087
1088	ASMAtomicAndU64(pu64, UINT64_C(0x8000000000000000));
1089	CHECKVAL(*pu64, UINT64_C(0x8000000000000000), "%x");
1090
1091	ASMAtomicAndU64(pu64, UINT64_C(0));
1092	CHECKVAL(*pu64, UINT64_C(0), "%x");
1093
1094	ASMAtomicOrU64(pu64, UINT64_C(0x4242424242424242));
1095	CHECKVAL(*pu64, UINT64_C(0x4242424242424242), "%x");
1096	}
1097
1098
1099	static void tstASMAtomicAndOrU64(void)
1100	{
1101	DO_SIMPLE_TEST(ASMAtomicAndOrU64, uint64_t);
1102	}
1103
1104
1105	typedef struct
1106	{
1107	uint8_t ab[PAGE_SIZE];
1108	} TSTPAGE;
1109
1110
1111	DECLINLINE(void) tstASMMemZeroPageWorker(TSTPAGE *pPage)
1112	{
1113	for (unsigned j = 0; j < 16; j++)
1114	{
1115	memset(pPage, 0x11 * j, sizeof(*pPage));
1116	ASMMemZeroPage(pPage);
1117	for (unsigned i = 0; i < sizeof(pPage->ab); i++)
1118	if (pPage->ab[i])
1119	RTTestFailed(g_hTest, "ASMMemZeroPage didn't clear byte at offset %#x!\n", i);
1120	}
1121	}
1122
1123
1124	static void tstASMMemZeroPage(void)
1125	{
1126	DO_SIMPLE_TEST(ASMMemZeroPage, TSTPAGE);
1127	}
1128
1129
1130	void tstASMMemIsZeroPage(RTTEST hTest)
1131	{
1132	RTTestSub(hTest, "ASMMemIsZeroPage");
1133
1134	void *pvPage1 = RTTestGuardedAllocHead(hTest, PAGE_SIZE);
1135	void *pvPage2 = RTTestGuardedAllocTail(hTest, PAGE_SIZE);
1136	RTTESTI_CHECK_RETV(pvPage1 && pvPage2);
1137
1138	memset(pvPage1, 0, PAGE_SIZE);
1139	memset(pvPage2, 0, PAGE_SIZE);
1140	RTTESTI_CHECK(ASMMemIsZeroPage(pvPage1));
1141	RTTESTI_CHECK(ASMMemIsZeroPage(pvPage2));
1142
1143	memset(pvPage1, 0xff, PAGE_SIZE);
1144	memset(pvPage2, 0xff, PAGE_SIZE);
1145	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage1));
1146	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage2));
1147
1148	memset(pvPage1, 0, PAGE_SIZE);
1149	memset(pvPage2, 0, PAGE_SIZE);
1150	for (unsigned off = 0; off < PAGE_SIZE; off++)
1151	{
1152	((uint8_t *)pvPage1)[off] = 1;
1153	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage1));
1154	((uint8_t *)pvPage1)[off] = 0;
1155
1156	((uint8_t *)pvPage2)[off] = 0x80;
1157	RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage2));
1158	((uint8_t *)pvPage2)[off] = 0;
1159	}
1160
1161	RTTestSubDone(hTest);
1162	}
1163
1164
1165	void tstASMMemZero32(void)
1166	{
1167	RTTestSub(g_hTest, "ASMMemFill32");
1168
1169	struct
1170	{
1171	uint64_t u64Magic1;
1172	uint8_t abPage[PAGE_SIZE - 32];
1173	uint64_t u64Magic2;
1174	} Buf1, Buf2, Buf3;
1175
1176	Buf1.u64Magic1 = UINT64_C(0xffffffffffffffff);
1177	memset(Buf1.abPage, 0x55, sizeof(Buf1.abPage));
1178	Buf1.u64Magic2 = UINT64_C(0xffffffffffffffff);
1179	Buf2.u64Magic1 = UINT64_C(0xffffffffffffffff);
1180	memset(Buf2.abPage, 0x77, sizeof(Buf2.abPage));
1181	Buf2.u64Magic2 = UINT64_C(0xffffffffffffffff);
1182	Buf3.u64Magic1 = UINT64_C(0xffffffffffffffff);
1183	memset(Buf3.abPage, 0x99, sizeof(Buf3.abPage));
1184	Buf3.u64Magic2 = UINT64_C(0xffffffffffffffff);
1185	ASMMemZero32(Buf1.abPage, sizeof(Buf1.abPage));
1186	ASMMemZero32(Buf2.abPage, sizeof(Buf2.abPage));
1187	ASMMemZero32(Buf3.abPage, sizeof(Buf3.abPage));
1188	if ( Buf1.u64Magic1 != UINT64_C(0xffffffffffffffff)
1189	\|\| Buf1.u64Magic2 != UINT64_C(0xffffffffffffffff)
1190	\|\| Buf2.u64Magic1 != UINT64_C(0xffffffffffffffff)
1191	\|\| Buf2.u64Magic2 != UINT64_C(0xffffffffffffffff)
1192	\|\| Buf3.u64Magic1 != UINT64_C(0xffffffffffffffff)
1193	\|\| Buf3.u64Magic2 != UINT64_C(0xffffffffffffffff))
1194	{
1195	RTTestFailed(g_hTest, "ASMMemZero32 violated one/both magic(s)!\n");
1196	}
1197	for (unsigned i = 0; i < RT_ELEMENTS(Buf1.abPage); i++)
1198	if (Buf1.abPage[i])
1199	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
1200	for (unsigned i = 0; i < RT_ELEMENTS(Buf2.abPage); i++)
1201	if (Buf2.abPage[i])
1202	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
1203	for (unsigned i = 0; i < RT_ELEMENTS(Buf3.abPage); i++)
1204	if (Buf3.abPage[i])
1205	RTTestFailed(g_hTest, "ASMMemZero32 didn't clear byte at offset %#x!\n", i);
1206	}
1207
1208
1209	void tstASMMemFill32(void)
1210	{
1211	RTTestSub(g_hTest, "ASMMemFill32");
1212
1213	struct
1214	{
1215	uint64_t u64Magic1;
1216	uint32_t au32Page[PAGE_SIZE / 4];
1217	uint64_t u64Magic2;
1218	} Buf1;
1219	struct
1220	{
1221	uint64_t u64Magic1;
1222	uint32_t au32Page[(PAGE_SIZE / 4) - 3];
1223	uint64_t u64Magic2;
1224	} Buf2;
1225	struct
1226	{
1227	uint64_t u64Magic1;
1228	uint32_t au32Page[(PAGE_SIZE / 4) - 1];
1229	uint64_t u64Magic2;
1230	} Buf3;
1231
1232	Buf1.u64Magic1 = UINT64_C(0xffffffffffffffff);
1233	memset(Buf1.au32Page, 0x55, sizeof(Buf1.au32Page));
1234	Buf1.u64Magic2 = UINT64_C(0xffffffffffffffff);
1235	Buf2.u64Magic1 = UINT64_C(0xffffffffffffffff);
1236	memset(Buf2.au32Page, 0x77, sizeof(Buf2.au32Page));
1237	Buf2.u64Magic2 = UINT64_C(0xffffffffffffffff);
1238	Buf3.u64Magic1 = UINT64_C(0xffffffffffffffff);
1239	memset(Buf3.au32Page, 0x99, sizeof(Buf3.au32Page));
1240	Buf3.u64Magic2 = UINT64_C(0xffffffffffffffff);
1241	ASMMemFill32(Buf1.au32Page, sizeof(Buf1.au32Page), 0xdeadbeef);
1242	ASMMemFill32(Buf2.au32Page, sizeof(Buf2.au32Page), 0xcafeff01);
1243	ASMMemFill32(Buf3.au32Page, sizeof(Buf3.au32Page), 0xf00dd00f);
1244	if ( Buf1.u64Magic1 != UINT64_C(0xffffffffffffffff)
1245	\|\| Buf1.u64Magic2 != UINT64_C(0xffffffffffffffff)
1246	\|\| Buf2.u64Magic1 != UINT64_C(0xffffffffffffffff)
1247	\|\| Buf2.u64Magic2 != UINT64_C(0xffffffffffffffff)
1248	\|\| Buf3.u64Magic1 != UINT64_C(0xffffffffffffffff)
1249	\|\| Buf3.u64Magic2 != UINT64_C(0xffffffffffffffff))
1250	RTTestFailed(g_hTest, "ASMMemFill32 violated one/both magic(s)!\n");
1251	for (unsigned i = 0; i < RT_ELEMENTS(Buf1.au32Page); i++)
1252	if (Buf1.au32Page[i] != 0xdeadbeef)
1253	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf1.au32Page[i], 0xdeadbeef);
1254	for (unsigned i = 0; i < RT_ELEMENTS(Buf2.au32Page); i++)
1255	if (Buf2.au32Page[i] != 0xcafeff01)
1256	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf2.au32Page[i], 0xcafeff01);
1257	for (unsigned i = 0; i < RT_ELEMENTS(Buf3.au32Page); i++)
1258	if (Buf3.au32Page[i] != 0xf00dd00f)
1259	RTTestFailed(g_hTest, "ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf3.au32Page[i], 0xf00dd00f);
1260	}
1261
1262
1263
1264	void tstASMMath(void)
1265	{
1266	RTTestSub(g_hTest, "Math");
1267
1268	uint64_t u64 = ASMMult2xU32RetU64(UINT32_C(0x80000000), UINT32_C(0x10000000));
1269	CHECKVAL(u64, UINT64_C(0x0800000000000000), "%#018RX64");
1270
1271	uint32_t u32 = ASMDivU64ByU32RetU32(UINT64_C(0x0800000000000000), UINT32_C(0x10000000));
1272	CHECKVAL(u32, UINT32_C(0x80000000), "%#010RX32");
1273
1274	#if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
1275	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x0000000000000001), UINT32_C(0x00000001), UINT32_C(0x00000001));
1276	CHECKVAL(u64, UINT64_C(0x0000000000000001), "%#018RX64");
1277	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x0000000100000000), UINT32_C(0x80000000), UINT32_C(0x00000002));
1278	CHECKVAL(u64, UINT64_C(0x4000000000000000), "%#018RX64");
1279	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xfedcba9876543210), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
1280	CHECKVAL(u64, UINT64_C(0xfedcba9876543210), "%#018RX64");
1281	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xffffffffffffffff), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
1282	CHECKVAL(u64, UINT64_C(0xffffffffffffffff), "%#018RX64");
1283	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xffffffffffffffff), UINT32_C(0xfffffff0), UINT32_C(0xffffffff));
1284	CHECKVAL(u64, UINT64_C(0xfffffff0fffffff0), "%#018RX64");
1285	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x3415934810359583), UINT32_C(0x58734981), UINT32_C(0xf8694045));
1286	CHECKVAL(u64, UINT64_C(0x128b9c3d43184763), "%#018RX64");
1287	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x3415934810359583), UINT32_C(0xf8694045), UINT32_C(0x58734981));
1288	CHECKVAL(u64, UINT64_C(0x924719355cd35a27), "%#018RX64");
1289
1290	# if 0 /* bird: question is whether this should trap or not:
1291	*
1292	* frank: Of course it must trap:
1293	*
1294	* 0xfffffff8 * 0x77d7daf8 = 0x77d7daf441412840
1295	*
1296	* During the following division, the quotient must fit into a 32-bit register.
1297	* Therefore the smallest valid divisor is
1298	*
1299	* (0x77d7daf441412840 >> 32) + 1 = 0x77d7daf5
1300	*
1301	* which is definitely greater than 0x3b9aca00.
1302	*
1303	* bird: No, the C version does not crash. So, the question is whether there's any
1304	* code depending on it not crashing.
1305	*
1306	* Of course the assembly versions of the code crash right now for the reasons you've
1307	* given, but the 32-bit MSC version does not crash.
1308	*
1309	* frank: The C version does not crash but delivers incorrect results for this case.
1310	* The reason is
1311	*
1312	* u.s.Hi = (unsigned long)(u64Hi / u32C);
1313	*
1314	* Here the division is actually 64-bit by 64-bit but the 64-bit result is truncated
1315	* to 32 bit. If using this (optimized and fast) function we should just be sure that
1316	* the operands are in a valid range.
1317	*/
1318	u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xfffffff8c65d6731), UINT32_C(0x77d7daf8), UINT32_C(0x3b9aca00));
1319	CHECKVAL(u64, UINT64_C(0x02b8f9a2aa74e3dc), "%#018RX64");
1320	# endif
1321	#endif /* AMD64 \|\| X86 */
1322
1323	u32 = ASMModU64ByU32RetU32(UINT64_C(0x0ffffff8c65d6731), UINT32_C(0x77d7daf8));
1324	CHECKVAL(u32, UINT32_C(0x3B642451), "%#010RX32");
1325
1326	int32_t i32;
1327	i32 = ASMModS64ByS32RetS32(INT64_C(-11), INT32_C(-2));
1328	CHECKVAL(i32, INT32_C(-1), "%010RI32");
1329	i32 = ASMModS64ByS32RetS32(INT64_C(-11), INT32_C(2));
1330	CHECKVAL(i32, INT32_C(-1), "%010RI32");
1331	i32 = ASMModS64ByS32RetS32(INT64_C(11), INT32_C(-2));
1332	CHECKVAL(i32, INT32_C(1), "%010RI32");
1333
1334	i32 = ASMModS64ByS32RetS32(INT64_C(92233720368547758), INT32_C(2147483647));
1335	CHECKVAL(i32, INT32_C(2104533974), "%010RI32");
1336	i32 = ASMModS64ByS32RetS32(INT64_C(-92233720368547758), INT32_C(2147483647));
1337	CHECKVAL(i32, INT32_C(-2104533974), "%010RI32");
1338	}
1339
1340
1341	void tstASMByteSwap(void)
1342	{
1343	RTTestSub(g_hTest, "ASMByteSwap*");
1344
1345	uint64_t u64In = UINT64_C(0x0011223344556677);
1346	uint64_t u64Out = ASMByteSwapU64(u64In);
1347	CHECKVAL(u64In, UINT64_C(0x0011223344556677), "%#018RX64");
1348	CHECKVAL(u64Out, UINT64_C(0x7766554433221100), "%#018RX64");
1349	u64Out = ASMByteSwapU64(u64Out);
1350	CHECKVAL(u64Out, u64In, "%#018RX64");
1351	u64In = UINT64_C(0x0123456789abcdef);
1352	u64Out = ASMByteSwapU64(u64In);
1353	CHECKVAL(u64In, UINT64_C(0x0123456789abcdef), "%#018RX64");
1354	CHECKVAL(u64Out, UINT64_C(0xefcdab8967452301), "%#018RX64");
1355	u64Out = ASMByteSwapU64(u64Out);
1356	CHECKVAL(u64Out, u64In, "%#018RX64");
1357	u64In = 0;
1358	u64Out = ASMByteSwapU64(u64In);
1359	CHECKVAL(u64Out, u64In, "%#018RX64");
1360	u64In = ~(uint64_t)0;
1361	u64Out = ASMByteSwapU64(u64In);
1362	CHECKVAL(u64Out, u64In, "%#018RX64");
1363
1364	uint32_t u32In = UINT32_C(0x00112233);
1365	uint32_t u32Out = ASMByteSwapU32(u32In);
1366	CHECKVAL(u32In, UINT32_C(0x00112233), "%#010RX32");
1367	CHECKVAL(u32Out, UINT32_C(0x33221100), "%#010RX32");
1368	u32Out = ASMByteSwapU32(u32Out);
1369	CHECKVAL(u32Out, u32In, "%#010RX32");
1370	u32In = UINT32_C(0x12345678);
1371	u32Out = ASMByteSwapU32(u32In);
1372	CHECKVAL(u32In, UINT32_C(0x12345678), "%#010RX32");
1373	CHECKVAL(u32Out, UINT32_C(0x78563412), "%#010RX32");
1374	u32Out = ASMByteSwapU32(u32Out);
1375	CHECKVAL(u32Out, u32In, "%#010RX32");
1376	u32In = 0;
1377	u32Out = ASMByteSwapU32(u32In);
1378	CHECKVAL(u32Out, u32In, "%#010RX32");
1379	u32In = ~(uint32_t)0;
1380	u32Out = ASMByteSwapU32(u32In);
1381	CHECKVAL(u32Out, u32In, "%#010RX32");
1382
1383	uint16_t u16In = UINT16_C(0x0011);
1384	uint16_t u16Out = ASMByteSwapU16(u16In);
1385	CHECKVAL(u16In, UINT16_C(0x0011), "%#06RX16");
1386	CHECKVAL(u16Out, UINT16_C(0x1100), "%#06RX16");
1387	u16Out = ASMByteSwapU16(u16Out);
1388	CHECKVAL(u16Out, u16In, "%#06RX16");
1389	u16In = UINT16_C(0x1234);
1390	u16Out = ASMByteSwapU16(u16In);
1391	CHECKVAL(u16In, UINT16_C(0x1234), "%#06RX16");
1392	CHECKVAL(u16Out, UINT16_C(0x3412), "%#06RX16");
1393	u16Out = ASMByteSwapU16(u16Out);
1394	CHECKVAL(u16Out, u16In, "%#06RX16");
1395	u16In = 0;
1396	u16Out = ASMByteSwapU16(u16In);
1397	CHECKVAL(u16Out, u16In, "%#06RX16");
1398	u16In = ~(uint16_t)0;
1399	u16Out = ASMByteSwapU16(u16In);
1400	CHECKVAL(u16Out, u16In, "%#06RX16");
1401	}
1402
1403
1404	void tstASMBench(void)
1405	{
1406	/*
1407	* Make this static. We don't want to have this located on the stack.
1408	*/
1409	static uint8_t volatile s_u8;
1410	static int8_t volatile s_i8;
1411	static uint16_t volatile s_u16;
1412	static int16_t volatile s_i16;
1413	static uint32_t volatile s_u32;
1414	static int32_t volatile s_i32;
1415	static uint64_t volatile s_u64;
1416	static int64_t volatile s_i64;
1417	register unsigned i;
1418	const unsigned cRounds = _2M;
1419	register uint64_t u64Elapsed;
1420
1421	RTTestSub(g_hTest, "Benchmarking");
1422
1423	#if 0 && !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
1424	# define BENCH(op, str) \
1425	do { \
1426	RTThreadYield(); \
1427	u64Elapsed = ASMReadTSC(); \
1428	for (i = cRounds; i > 0; i--) \
1429	op; \
1430	u64Elapsed = ASMReadTSC() - u64Elapsed; \
1431	RTTestValue(g_hTest, str, u64Elapsed / cRounds, RTTESTUNIT_TICKS_PER_CALL); \
1432	} while (0)
1433	#else
1434	# define BENCH(op, str) \
1435	do { \
1436	RTThreadYield(); \
1437	u64Elapsed = RTTimeNanoTS(); \
1438	for (i = cRounds; i > 0; i--) \
1439	op; \
1440	u64Elapsed = RTTimeNanoTS() - u64Elapsed; \
1441	RTTestValue(g_hTest, str, u64Elapsed / cRounds, RTTESTUNIT_NS_PER_CALL); \
1442	} while (0)
1443	#endif
1444
1445	BENCH(s_u32 = 0, "s_u32 = 0");
1446	BENCH(ASMAtomicUoReadU8(&s_u8), "ASMAtomicUoReadU8");
1447	BENCH(ASMAtomicUoReadS8(&s_i8), "ASMAtomicUoReadS8");
1448	BENCH(ASMAtomicUoReadU16(&s_u16), "ASMAtomicUoReadU16");
1449	BENCH(ASMAtomicUoReadS16(&s_i16), "ASMAtomicUoReadS16");
1450	BENCH(ASMAtomicUoReadU32(&s_u32), "ASMAtomicUoReadU32");
1451	BENCH(ASMAtomicUoReadS32(&s_i32), "ASMAtomicUoReadS32");
1452	BENCH(ASMAtomicUoReadU64(&s_u64), "ASMAtomicUoReadU64");
1453	BENCH(ASMAtomicUoReadS64(&s_i64), "ASMAtomicUoReadS64");
1454	BENCH(ASMAtomicReadU8(&s_u8), "ASMAtomicReadU8");
1455	BENCH(ASMAtomicReadS8(&s_i8), "ASMAtomicReadS8");
1456	BENCH(ASMAtomicReadU16(&s_u16), "ASMAtomicReadU16");
1457	BENCH(ASMAtomicReadS16(&s_i16), "ASMAtomicReadS16");
1458	BENCH(ASMAtomicReadU32(&s_u32), "ASMAtomicReadU32");
1459	BENCH(ASMAtomicReadS32(&s_i32), "ASMAtomicReadS32");
1460	BENCH(ASMAtomicReadU64(&s_u64), "ASMAtomicReadU64");
1461	BENCH(ASMAtomicReadS64(&s_i64), "ASMAtomicReadS64");
1462	BENCH(ASMAtomicUoWriteU8(&s_u8, 0), "ASMAtomicUoWriteU8");
1463	BENCH(ASMAtomicUoWriteS8(&s_i8, 0), "ASMAtomicUoWriteS8");
1464	BENCH(ASMAtomicUoWriteU16(&s_u16, 0), "ASMAtomicUoWriteU16");
1465	BENCH(ASMAtomicUoWriteS16(&s_i16, 0), "ASMAtomicUoWriteS16");
1466	BENCH(ASMAtomicUoWriteU32(&s_u32, 0), "ASMAtomicUoWriteU32");
1467	BENCH(ASMAtomicUoWriteS32(&s_i32, 0), "ASMAtomicUoWriteS32");
1468	BENCH(ASMAtomicUoWriteU64(&s_u64, 0), "ASMAtomicUoWriteU64");
1469	BENCH(ASMAtomicUoWriteS64(&s_i64, 0), "ASMAtomicUoWriteS64");
1470	BENCH(ASMAtomicWriteU8(&s_u8, 0), "ASMAtomicWriteU8");
1471	BENCH(ASMAtomicWriteS8(&s_i8, 0), "ASMAtomicWriteS8");
1472	BENCH(ASMAtomicWriteU16(&s_u16, 0), "ASMAtomicWriteU16");
1473	BENCH(ASMAtomicWriteS16(&s_i16, 0), "ASMAtomicWriteS16");
1474	BENCH(ASMAtomicWriteU32(&s_u32, 0), "ASMAtomicWriteU32");
1475	BENCH(ASMAtomicWriteS32(&s_i32, 0), "ASMAtomicWriteS32");
1476	BENCH(ASMAtomicWriteU64(&s_u64, 0), "ASMAtomicWriteU64");
1477	BENCH(ASMAtomicWriteS64(&s_i64, 0), "ASMAtomicWriteS64");
1478	BENCH(ASMAtomicXchgU8(&s_u8, 0), "ASMAtomicXchgU8");
1479	BENCH(ASMAtomicXchgS8(&s_i8, 0), "ASMAtomicXchgS8");
1480	BENCH(ASMAtomicXchgU16(&s_u16, 0), "ASMAtomicXchgU16");
1481	BENCH(ASMAtomicXchgS16(&s_i16, 0), "ASMAtomicXchgS16");
1482	BENCH(ASMAtomicXchgU32(&s_u32, 0), "ASMAtomicXchgU32");
1483	BENCH(ASMAtomicXchgS32(&s_i32, 0), "ASMAtomicXchgS32");
1484	BENCH(ASMAtomicXchgU64(&s_u64, 0), "ASMAtomicXchgU64");
1485	BENCH(ASMAtomicXchgS64(&s_i64, 0), "ASMAtomicXchgS64");
1486	BENCH(ASMAtomicCmpXchgU32(&s_u32, 0, 0), "ASMAtomicCmpXchgU32");
1487	BENCH(ASMAtomicCmpXchgS32(&s_i32, 0, 0), "ASMAtomicCmpXchgS32");
1488	BENCH(ASMAtomicCmpXchgU64(&s_u64, 0, 0), "ASMAtomicCmpXchgU64");
1489	BENCH(ASMAtomicCmpXchgS64(&s_i64, 0, 0), "ASMAtomicCmpXchgS64");
1490	BENCH(ASMAtomicCmpXchgU32(&s_u32, 0, 1), "ASMAtomicCmpXchgU32/neg");
1491	BENCH(ASMAtomicCmpXchgS32(&s_i32, 0, 1), "ASMAtomicCmpXchgS32/neg");
1492	BENCH(ASMAtomicCmpXchgU64(&s_u64, 0, 1), "ASMAtomicCmpXchgU64/neg");
1493	BENCH(ASMAtomicCmpXchgS64(&s_i64, 0, 1), "ASMAtomicCmpXchgS64/neg");
1494	BENCH(ASMAtomicIncU32(&s_u32), "ASMAtomicIncU32");
1495	BENCH(ASMAtomicIncS32(&s_i32), "ASMAtomicIncS32");
1496	BENCH(ASMAtomicDecU32(&s_u32), "ASMAtomicDecU32");
1497	BENCH(ASMAtomicDecS32(&s_i32), "ASMAtomicDecS32");
1498	BENCH(ASMAtomicAddU32(&s_u32, 5), "ASMAtomicAddU32");
1499	BENCH(ASMAtomicAddS32(&s_i32, 5), "ASMAtomicAddS32");
1500	/* The Darwin gcc does not like this ... */
1501	#if !defined(RT_OS_DARWIN) && !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
1502	BENCH(s_u8 = ASMGetApicId(), "ASMGetApicId");
1503	#endif
1504
1505	#undef BENCH
1506	}
1507
1508
1509	int main(int argc, char *argv[])
1510	{
1511	int rc = RTTestInitAndCreate("tstRTInlineAsm", &g_hTest);
1512	if (rc)
1513	return rc;
1514	RTTestBanner(g_hTest);
1515
1516	/*
1517	* Execute the tests.
1518	*/
1519	#if !defined(GCC44_32BIT_PIC) && (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86))
1520	tstASMCpuId();
1521	#endif
1522	tstASMAtomicXchgU8();
1523	tstASMAtomicXchgU16();
1524	tstASMAtomicXchgU32();
1525	tstASMAtomicXchgU64();
1526	tstASMAtomicXchgPtr();
1527	tstASMAtomicCmpXchgU8();
1528	tstASMAtomicCmpXchgU32();
1529	tstASMAtomicCmpXchgU64();
1530	tstASMAtomicCmpXchgExU32();
1531	tstASMAtomicCmpXchgExU64();
1532	tstASMAtomicReadU64();
1533	tstASMAtomicUoReadU64();
1534
1535	tstASMAtomicAddS32();
1536	tstASMAtomicAddS64();
1537	tstASMAtomicDecIncS32();
1538	tstASMAtomicDecIncS64();
1539	tstASMAtomicAndOrU32();
1540	tstASMAtomicAndOrU64();
1541
1542	tstASMMemZeroPage();
1543	tstASMMemIsZeroPage(g_hTest);
1544	tstASMMemZero32();
1545	tstASMMemFill32();
1546
1547	tstASMMath();
1548
1549	tstASMByteSwap();
1550
1551	tstASMBench();
1552
1553	/*
1554	* Show the result.
1555	*/
1556	return RTTestSummaryAndDestroy(g_hTest);
1557	}
1558

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

source: vbox/trunk/src/VBox/Runtime/testcase/tstRTInlineAsm.cpp@ 36363

Download in other formats: