bs3-cpu-instr-2-template.c@ 105686

Last change on this file since 105686 was 104066, checked in by vboxsync, 8 months ago
ValKit/bootsectors: Added a g_cBs3ThresholdNativeRecompiler variable that can be used instead of BS3_THRESHOLD_NATIVE_RECOMPILER, initializing it with the VMMDEV_TESTING_CFG_THRESHOLD_NATIVE_RECOMPILER value from the host if available. bugref:9898
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 245.6 KB

Line
1	/* $Id: bs3-cpu-instr-2-template.c 104066 2024-03-26 16:06:07Z vboxsync $ */
2	/** @file
3	* BS3Kit - bs3-cpu-instr-2, C code template.
4	*/
5
6	/*
7	* Copyright (C) 2007-2023 Oracle and/or its affiliates.
8	*
9	* This file is part of VirtualBox base platform packages, as
10	* available from https://www.virtualbox.org.
11	*
12	* This program is free software; you can redistribute it and/or
13	* modify it under the terms of the GNU General Public License
14	* as published by the Free Software Foundation, in version 3 of the
15	* License.
16	*
17	* This program is distributed in the hope that it will be useful, but
18	* WITHOUT ANY WARRANTY; without even the implied warranty of
19	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20	* General Public License for more details.
21	*
22	* You should have received a copy of the GNU General Public License
23	* along with this program; if not, see <https://www.gnu.org/licenses>.
24	*
25	* The contents of this file may alternatively be used under the terms
26	* of the Common Development and Distribution License Version 1.0
27	* (CDDL), a copy of it is provided in the "COPYING.CDDL" file included
28	* in the VirtualBox distribution, in which case the provisions of the
29	* CDDL are applicable instead of those of the GPL.
30	*
31	* You may elect to license modified versions of this file under the
32	* terms and conditions of either the GPL or the CDDL or both.
33	*
34	* SPDX-License-Identifier: GPL-3.0-only OR CDDL-1.0
35	*/
36
37
38	/*********************************************************************************************************************************
39	* Header Files *
40	*********************************************************************************************************************************/
41	#include <iprt/asm.h>
42	#include <iprt/asm-amd64-x86.h>
43	#include "bs3-cpu-instr-2.h"
44	#include "bs3-cpu-instr-2-data.h"
45	#include "bs3-cpu-instr-2-asm-auto.h"
46
47
48	/*********************************************************************************************************************************
49	* Structures and Typedefs *
50	*********************************************************************************************************************************/
51	#ifdef BS3_INSTANTIATING_CMN
52	# if ARCH_BITS == 64
53	typedef struct BS3CI2FSGSBASE
54	{
55	const char *pszDesc;
56	bool f64BitOperand;
57	FPFNBS3FAR pfnWorker;
58	uint8_t offWorkerUd2;
59	FPFNBS3FAR pfnVerifyWorker;
60	uint8_t offVerifyWorkerUd2;
61	} BS3CI2FSGSBASE;
62	# endif
63	#endif
64
65
66	/*********************************************************************************************************************************
67	* Global Variables *
68	*********************************************************************************************************************************/
69	#ifdef BS3_INSTANTIATING_CMN
70	# if ARCH_BITS == 64
71	static BS3CI2FSGSBASE const s_aWrFsBaseWorkers[] =
72	{
73	{ "wrfsbase rbx", true, BS3_CMN_NM(bs3CpuInstr2_wrfsbase_rbx_ud2), 5, BS3_CMN_NM(bs3CpuInstr2_wrfsbase_rbx_rdfsbase_rcx_ud2), 15 },
74	{ "wrfsbase ebx", false, BS3_CMN_NM(bs3CpuInstr2_wrfsbase_ebx_ud2), 4, BS3_CMN_NM(bs3CpuInstr2_wrfsbase_ebx_rdfsbase_ecx_ud2), 13 },
75	};
76
77	static BS3CI2FSGSBASE const s_aWrGsBaseWorkers[] =
78	{
79	{ "wrgsbase rbx", true, BS3_CMN_NM(bs3CpuInstr2_wrgsbase_rbx_ud2), 5, BS3_CMN_NM(bs3CpuInstr2_wrgsbase_rbx_rdgsbase_rcx_ud2), 15 },
80	{ "wrgsbase ebx", false, BS3_CMN_NM(bs3CpuInstr2_wrgsbase_ebx_ud2), 4, BS3_CMN_NM(bs3CpuInstr2_wrgsbase_ebx_rdgsbase_ecx_ud2), 13 },
81	};
82
83	static BS3CI2FSGSBASE const s_aRdFsBaseWorkers[] =
84	{
85	{ "rdfsbase rbx", true, BS3_CMN_NM(bs3CpuInstr2_rdfsbase_rbx_ud2), 5, BS3_CMN_NM(bs3CpuInstr2_wrfsbase_rbx_rdfsbase_rcx_ud2), 15 },
86	{ "rdfsbase ebx", false, BS3_CMN_NM(bs3CpuInstr2_rdfsbase_ebx_ud2), 4, BS3_CMN_NM(bs3CpuInstr2_wrfsbase_ebx_rdfsbase_ecx_ud2), 13 },
87	};
88
89	static BS3CI2FSGSBASE const s_aRdGsBaseWorkers[] =
90	{
91	{ "rdgsbase rbx", true, BS3_CMN_NM(bs3CpuInstr2_rdgsbase_rbx_ud2), 5, BS3_CMN_NM(bs3CpuInstr2_wrgsbase_rbx_rdgsbase_rcx_ud2), 15 },
92	{ "rdgsbase ebx", false, BS3_CMN_NM(bs3CpuInstr2_rdgsbase_ebx_ud2), 4, BS3_CMN_NM(bs3CpuInstr2_wrgsbase_ebx_rdgsbase_ecx_ud2), 13 },
93	};
94	# endif
95	#endif /* BS3_INSTANTIATING_CMN - global */
96
97
98	/*
99	* Common code.
100	* Common code.
101	* Common code.
102	*/
103	#ifdef BS3_INSTANTIATING_CMN
104
105	/*
106	* Basic binary arithmetic tests.
107	*/
108
109	# if ARCH_BITS == 64 /* fDstMem cBitsImm */
110	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_8_64BIT(a_Ins) \
111	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _sil_dil), X86_GREG_xSI, X86_GREG_xDI, false, false, 0 }, \
112	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r9b_r8b), X86_GREG_x9, X86_GREG_x8, false, false, 0 }, \
113	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _al_r13b), X86_GREG_xAX, X86_GREG_x13, false, false, 0 }, \
114	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSx14_r11b), X86_GREG_x14, X86_GREG_x11, true, false, 0 },
115	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8_64BIT(a_Ins) \
116	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8b_Ib), X86_GREG_x8, X86_GREG_x15, false, false, 8 }, \
117	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r14b_Ib), X86_GREG_x14, X86_GREG_x15, false, false, 8 }, \
118	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSx13_Ib), X86_GREG_x13, X86_GREG_x15, true, false, 8 },
119	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8_64BIT(a_Ins) \
120	{ BS3_CMN_NM(bs3CpuInstr2_alt_ ## a_Ins ## _dl_r14b), X86_GREG_xDX, X86_GREG_x14, false, false, 0 }, \
121	{ BS3_CMN_NM(bs3CpuInstr2_alt_ ## a_Ins ## _r8b_bl), X86_GREG_x8, X86_GREG_xBX, false, false, 0 }, \
122	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r11b_DSx12), X86_GREG_x11, X86_GREG_x12, false, true, 0 },
123	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_16_64BIT(a_Ins) \
124	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8w_cx), X86_GREG_x8, X86_GREG_xCX, false, false, 0 }, \
125	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r15w_r10w), X86_GREG_x15, X86_GREG_x10, false, false, 0 }, \
126	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSx15_r12w), X86_GREG_x15, X86_GREG_x12, true, false, 0 },
127	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B_64BIT(a_Ins) \
128	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8w_Ib), X86_GREG_x8, X86_GREG_xBX, false, false, 8 }, \
129	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r12w_Ib), X86_GREG_x12, X86_GREG_xBX, false, false, 8 }, \
130	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _wDSx14_Ib), X86_GREG_x14, X86_GREG_xBX, true, false, 8 },
131	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W_64BIT(a_Ins) \
132	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8w_Iw), X86_GREG_x8, X86_GREG_xBX, false, false, 16 }, \
133	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r13w_Iw), X86_GREG_x13, X86_GREG_xBX, false, false, 16 }, \
134	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSx11_Iw), X86_GREG_x11, X86_GREG_xBX, true, false, 16 },
135	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16_64BIT(a_Ins) \
136	{ BS3_CMN_NM(bs3CpuInstr2_alt_ ## a_Ins ## _r13w_ax), X86_GREG_x13, X86_GREG_xAX, false, false, 0 }, \
137	{ BS3_CMN_NM(bs3CpuInstr2_alt_ ## a_Ins ## _si_r9w), X86_GREG_xSI, X86_GREG_x9, false, false, 0 }, \
138	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r9w_DSx8), X86_GREG_x9, X86_GREG_x8, false, true, 0 },
139	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_32_64BIT(a_Ins) \
140	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _eax_r8d), X86_GREG_xAX, X86_GREG_x8, false, false, 0 }, \
141	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r9d_ecx), X86_GREG_x9, X86_GREG_xCX, false, false, 0 }, \
142	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r13d_r14d), X86_GREG_x13, X86_GREG_x14, false, false, 0 }, \
143	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSx10_r11d), X86_GREG_x10, X86_GREG_x11, true, false, 0 },
144	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B_64BIT(a_Ins) \
145	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8d_Ib), X86_GREG_x8, X86_GREG_xBX, false, false, 8 }, \
146	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r11d_Ib), X86_GREG_x11, X86_GREG_xBX, false, false, 8 }, \
147	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dwDSx15_Ib), X86_GREG_x15, X86_GREG_xBX, true, false, 8 },
148	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW_64BIT(a_Ins) \
149	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8d_Id), X86_GREG_x8, X86_GREG_xBX, false, false, 32 }, \
150	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r14d_Id), X86_GREG_x14, X86_GREG_xBX, false, false, 32 }, \
151	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSx12_Id), X86_GREG_x12, X86_GREG_xBX, true, false, 32 },
152	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_32_64BIT(a_Ins) \
153	{ BS3_CMN_NM(bs3CpuInstr2_alt_ ## a_Ins ## _r15d_esi), X86_GREG_x15, X86_GREG_xSI, false, false, 0 }, \
154	{ BS3_CMN_NM(bs3CpuInstr2_alt_ ## a_Ins ## _eax_r10d), X86_GREG_xAX, X86_GREG_x10, false, false, 0 }, \
155	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r14d_DSx12), X86_GREG_x14, X86_GREG_x12, false, true, 0 },
156
157	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_64(a_Ins) \
158	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _rax_rbx), X86_GREG_xAX, X86_GREG_xBX, false, false, 0 }, \
159	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8_rax), X86_GREG_x8, X86_GREG_xAX, false, false, 0 }, \
160	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _rdx_r10), X86_GREG_xDX, X86_GREG_x10, false, false, 0 }, \
161	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSxBX_rax), X86_GREG_xBX, X86_GREG_xAX, true, false, 0 }, \
162	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSx12_r8), X86_GREG_x12, X86_GREG_x8, true, false, 0 },
163	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_64(a_Ins) \
164	BS3CPUINSTR2CMNBINTEST_ENTRIES_64(a_Ins) \
165	{ BS3_CMN_NM(bs3CpuInstr2_alt_ ## a_Ins ## _r15_rsi), X86_GREG_x15, X86_GREG_xSI, false, false, 0 }, \
166	{ BS3_CMN_NM(bs3CpuInstr2_alt_ ## a_Ins ## _rbx_r14), X86_GREG_xBX, X86_GREG_x14, false, false, 0 }, \
167	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _rax_DSxBX), X86_GREG_xAX, X86_GREG_xBX, false, true, 0 }, \
168	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8_DSx12), X86_GREG_x8, X86_GREG_x12, false, true, 0 },
169
170	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(a_Ins) \
171	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _rax_Ib), X86_GREG_xAX, X86_GREG_x15, false, false, 8 }, \
172	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _rbp_Ib), X86_GREG_xBP, X86_GREG_x15, false, false, 8 }, \
173	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8_Ib), X86_GREG_x8, X86_GREG_x15, false, false, 8 }, \
174	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r11_Ib), X86_GREG_x11, X86_GREG_x15, false, false, 8 }, \
175	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _qwDSxSI_Ib), X86_GREG_xSI, X86_GREG_x15, true, false, 8 }, \
176	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _qwDSx8_Ib), X86_GREG_x8, X86_GREG_x15, true, false, 8 },
177
178	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64DW(a_Ins) \
179	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _rax_Id), X86_GREG_xAX, X86_GREG_x15, false, false, 32 }, \
180	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8_Id), X86_GREG_x8, X86_GREG_x15, false, false, 32 }, \
181	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _rbx_Id), X86_GREG_xBX, X86_GREG_x15, false, false, 32 }, \
182	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r14_Id), X86_GREG_x14, X86_GREG_x15, false, false, 32 }, \
183	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _qwDSx12_Id), X86_GREG_x12, X86_GREG_x15, true, false, 32 },
184
185	# else
186	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_8_64BIT(aIns)
187	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_16_64BIT(aIns)
188	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_32_64BIT(aIns)
189	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8_64BIT(aIns)
190	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B_64BIT(aIns)
191	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W_64BIT(aIns)
192	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B_64BIT(aIns)
193	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW_64BIT(aIns)
194	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8_64BIT(aIns)
195	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16_64BIT(aIns)
196	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_32_64BIT(aIns)
197	# endif
198
199	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_8(a_Ins) \
200	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _al_dl), X86_GREG_xAX, X86_GREG_xDX, false, false, 0 }, \
201	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ch_bh), X86_GREG_xCX+16, X86_GREG_xBX+16, false, false, 0 }, \
202	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dl_ah), X86_GREG_xDX, X86_GREG_xAX+16, false, false, 0 }, \
203	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSxBX_ah), X86_GREG_xBX, X86_GREG_xAX+16, true, false, 0 }, \
204	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSxDI_bl), X86_GREG_xDI, X86_GREG_xBX, true, false, 0 }, \
205	BS3CPUINSTR2CMNBINTEST_ENTRIES_8_64BIT(a_Ins)
206	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8(a_Ins) \
207	BS3CPUINSTR2CMNBINTEST_ENTRIES_8(a_Ins) \
208	{ BS3_CMN_NM(bs3CpuInstr2_alt_ ## a_Ins ## _dh_cl), X86_GREG_xDX+16, X86_GREG_xCX, false, false, 0 }, \
209	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dl_DSxBX), X86_GREG_xDX, X86_GREG_xBX, false, true, 0 }, \
210	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ch_DSxBX), X86_GREG_xCX+16, X86_GREG_xBX, false, true, 0 }, \
211	BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8_64BIT(a_Ins)
212
213	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8(a_Ins) \
214	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _al_Ib), X86_GREG_xAX, X86_GREG_x15, false, false, 8 }, \
215	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _cl_Ib), X86_GREG_xCX, X86_GREG_x15, false, false, 8 }, \
216	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dh_Ib), X86_GREG_xDX+16, X86_GREG_x15, false, false, 8 }, \
217	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSxDI_Ib), X86_GREG_xDI, X86_GREG_x15, true, false, 8 }, \
218	BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8_64BIT(a_Ins)
219
220	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_16(a_Ins) \
221	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _di_si), X86_GREG_xDI, X86_GREG_xSI, false, false, 0 }, \
222	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _cx_bp), X86_GREG_xCX, X86_GREG_xBP, false, false, 0 }, \
223	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSxDI_si), X86_GREG_xDI, X86_GREG_xSI, true, false, 0 }, \
224	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSxBX_ax), X86_GREG_xBX, X86_GREG_xAX, true, false, 0 }, \
225	BS3CPUINSTR2CMNBINTEST_ENTRIES_16_64BIT(a_Ins)
226	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16(a_Ins) \
227	BS3CPUINSTR2CMNBINTEST_ENTRIES_16(a_Ins) \
228	{ BS3_CMN_NM(bs3CpuInstr2_alt_ ## a_Ins ## _bp_bx), X86_GREG_xBP, X86_GREG_xBX, false, false, 0 }, \
229	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _si_DSxDI), X86_GREG_xSI, X86_GREG_xDI, false, true, 0 }, \
230	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ax_DSxBX), X86_GREG_xAX, X86_GREG_xBX, false, true, 0 }, \
231	BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16_64BIT(a_Ins)
232
233	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(a_Ins) \
234	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ax_Ib), X86_GREG_xAX, X86_GREG_x15, false, false, 8 }, \
235	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _si_Ib), X86_GREG_xSI, X86_GREG_x15, false, false, 8 }, \
236	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _wDSxDI_Ib), X86_GREG_xDI, X86_GREG_x15, true, false, 8 }, \
237	BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B_64BIT(a_Ins)
238
239	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W(a_Ins) \
240	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ax_Iw), X86_GREG_xAX, X86_GREG_x15, false, false, 16 }, \
241	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _bx_Iw), X86_GREG_xBX, X86_GREG_x15, false, false, 16 }, \
242	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSxBX_Iw), X86_GREG_xBX, X86_GREG_x15, true, false, 16 }, \
243	BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W_64BIT(a_Ins)
244
245
246	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_32(a_Ins) \
247	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _eax_ebx), X86_GREG_xAX, X86_GREG_xBX, false, false, 0 }, \
248	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ecx_ebp), X86_GREG_xCX, X86_GREG_xBP, false, false, 0 }, \
249	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _edx_edi), X86_GREG_xDX, X86_GREG_xDI, false, false, 0 }, \
250	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSxDI_esi), X86_GREG_xDI, X86_GREG_xSI, true, false, 0 }, \
251	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSxBX_eax), X86_GREG_xBX, X86_GREG_xAX, true, false, 0 }, \
252	BS3CPUINSTR2CMNBINTEST_ENTRIES_32_64BIT(a_Ins)
253	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_32(a_Ins) \
254	BS3CPUINSTR2CMNBINTEST_ENTRIES_32(a_Ins) \
255	{ BS3_CMN_NM(bs3CpuInstr2_alt_ ## a_Ins ## _edi_esi), X86_GREG_xDI, X86_GREG_xSI, false, false, 0 }, \
256	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _eax_DSxBX), X86_GREG_xAX, X86_GREG_xBX, false, true, 0 }, \
257	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ebp_DSxDI), X86_GREG_xBP, X86_GREG_xDI, false, true, 0 }, \
258	BS3CPUINSTR2CMNBINTEST_ENTRIES_32_64BIT(a_Ins)
259
260	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(a_Ins) \
261	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _eax_Ib), X86_GREG_xAX, X86_GREG_x15, false, false, 8 }, \
262	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ecx_Ib), X86_GREG_xCX, X86_GREG_x15, false, false, 8 }, \
263	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dwDSxDI_Ib), X86_GREG_xDI, X86_GREG_x15, true, false, 8 }, \
264	BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B_64BIT(a_Ins)
265
266	# define BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW(a_Ins) \
267	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _eax_Id), X86_GREG_xAX, X86_GREG_x15, false, false, 32 }, \
268	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ebp_Id), X86_GREG_xBP, X86_GREG_x15, false, false, 32 }, \
269	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _DSxSI_Id), X86_GREG_xSI, X86_GREG_x15, true, false, 32 }, \
270	BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW_64BIT(a_Ins)
271
272
273	typedef struct BS3CPUINSTR2CMNBINTEST
274	{
275	FPFNBS3FAR pfnWorker;
276	uint8_t idxDstReg;
277	uint8_t idxSrcReg;
278	bool fDstMem : 1;
279	bool fSrcMem : 1;
280	uint8_t cBitsImm;
281	} BS3CPUINSTR2CMNBINTEST;
282	typedef BS3CPUINSTR2CMNBINTEST const BS3_FAR_DATA *PCBS3CPUINSTR2CMNBINTEST;
283
284
285	static uint16_t const g_auEflStatusBitsVars[] =
286	{
287	0,
288	X86_EFL_STATUS_BITS,
289	X86_EFL_CF,
290	X86_EFL_PF,
291	X86_EFL_AF,
292	X86_EFL_ZF,
293	X86_EFL_SF,
294	X86_EFL_OF,
295	X86_EFL_PF \| X86_EFL_AF,
296	};
297
298
299	DECLINLINE(void RT_FAR ) Code2RwPtr(void RT_FAR pfn)
300	{
301	#if ARCH_BITS == 16
302	if (!BS3_MODE_IS_RM_OR_V86(g_bBs3CurrentMode))
303	pfn = BS3_FP_MAKE(BS3_SEL_TILED + 8, BS3_FP_OFF(pfn)); /* ASSUMES CS */
304	#endif
305	return pfn;
306	}
307
308	#define BS3CPUINSTR2_COMMON_BINARY_U(a_cBits, a_UIntType, a_UIntImmType, a_szFmt) \
309	static uint8_t \
310	RT_CONCAT(bs3CpuInstr2_CommonBinaryU,a_cBits)(uint8_t bMode, PCBS3CPUINSTR2CMNBINTEST paTests, unsigned cTests, uint16_t fPassthruEfl, \
311	RT_CONCAT(PCBS3CPUINSTR2BIN,a_cBits) paTestData, unsigned cTestData, bool fCarryIn, \
312	bool fMaskSrcWhenMemDst, bool fReadOnly) \
313	{ \
314	BS3REGCTX Ctx; \
315	BS3REGCTX CtxExpect; \
316	BS3TRAPFRAME TrapFrame; \
317	unsigned iTest; \
318	struct \
319	{ \
320	char achPreGuard[8]; \
321	a_UIntType uData; \
322	char achPostGuard[8]; \
323	} Buf = { { '0','1','2','3','4','5','6','7' }, 0, { '8','9','a','b','c','d','e','f'} }; \
324	a_UIntType uMemExpect = 0; \
325	a_UIntType uMemDummy = 0; \
326	\
327	/* Ensure the structures are allocated before we sample the stack pointer. */ \
328	Bs3MemSet(&Ctx, 0, sizeof(Ctx)); \
329	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame)); \
330	\
331	/* \
332	* Create test context. \
333	*/ \
334	Bs3RegCtxSaveEx(&Ctx, bMode, 640); \
335	Ctx.rflags.u32 &= ~X86_EFL_RF; \
336	if (ARCH_BITS == 64) \
337	for (iTest = 0; iTest < 16; iTest++) \
338	if (iTest != X86_GREG_xSP) \
339	(&Ctx.rax)[iTest].au32[1] = UINT32_C(0x8572ade) << (iTest & 7); \
340	Bs3MemCpy(&CtxExpect, &Ctx, sizeof(CtxExpect)); \
341	if (!BS3_MODE_IS_16BIT_SYS(bMode)) \
342	CtxExpect.rflags.u32 \|= X86_EFL_RF; \
343	\
344	/* \
345	* Each test worker. \
346	*/ \
347	for (iTest = 0; iTest < cTests; iTest++) \
348	{ \
349	uint8_t const cbInstr = ((uint8_t BS3_FAR )paTests[iTest].pfnWorker)[-1]; / the function is prefixed by the length */ \
350	uint8_t RT_FAR * const pbImm = (uint8_t BS3_FAR )Code2RwPtr(&((uint8_t BS3_FAR )paTests[iTest].pfnWorker)[cbInstr - 1]); \
351	a_UIntImmType RT_FAR * const puImm = (a_UIntImmType RT_FAR )Code2RwPtr(&((uint8_t BS3_FAR )paTests[iTest].pfnWorker)[cbInstr - sizeof(a_UIntImmType)]); \
352	unsigned const idxDstReg = paTests[iTest].idxDstReg; \
353	unsigned const idxSrcReg = paTests[iTest].idxSrcReg; \
354	uint16_t const SavedDs = Ctx.ds; \
355	BS3REG const SavedDst = (&Ctx.rax)[idxDstReg & 15]; /* saves memptr too */ \
356	BS3REG const SavedSrc = (&Ctx.rax)[idxSrcReg & 15]; /* ditto */ \
357	a_UIntType RT_FAR * const puCtxDst = paTests[iTest].fDstMem ? &Buf.uData \
358	: &(&Ctx.rax)[idxDstReg & 15].RT_CONCAT(au,a_cBits)[idxDstReg >> 4]; \
359	a_UIntType RT_FAR * const puCtxSrc = paTests[iTest].fSrcMem ? &Buf.uData \
360	: paTests[iTest].cBitsImm == 0 \
361	? &(&Ctx.rax)[idxSrcReg & 15].RT_CONCAT(au,a_cBits)[idxSrcReg >> 4] \
362	: &uMemDummy; \
363	a_UIntType RT_FAR * const puCtxExpectDst = paTests[iTest].fDstMem ? &uMemExpect \
364	: &(&CtxExpect.rax)[idxDstReg & 15].RT_CONCAT(au,a_cBits)[idxDstReg >> 4]; \
365	a_UIntType RT_FAR * const puCtxExpectSrc = paTests[iTest].fSrcMem ? &uMemExpect \
366	: paTests[iTest].cBitsImm == 0 \
367	? &(&CtxExpect.rax)[idxSrcReg & 15].RT_CONCAT(au,a_cBits)[idxSrcReg >> 4] \
368	: &uMemDummy; \
369	uint64_t RT_FAR * const puMemPtrReg = paTests[iTest].fDstMem ? &(&Ctx.rax)[idxDstReg & 15].u \
370	: paTests[iTest].fSrcMem ? &(&Ctx.rax)[idxSrcReg & 15].u : NULL; \
371	uint64_t RT_FAR * const puMemPtrRegExpt= paTests[iTest].fDstMem ? &(&CtxExpect.rax)[idxDstReg & 15].u \
372	: paTests[iTest].fSrcMem ? &(&CtxExpect.rax)[idxSrcReg & 15].u : NULL; \
373	unsigned iTestData; \
374	/Bs3TestPrintf("pfnWorker=%p\n", paTests[iTest].pfnWorker);/ \
375	\
376	Bs3RegCtxSetRipCsFromLnkPtr(&Ctx, paTests[iTest].pfnWorker); \
377	CtxExpect.rip.u = Ctx.rip.u + cbInstr; \
378	CtxExpect.cs = Ctx.cs; \
379	\
380	if (puMemPtrReg) \
381	CtxExpect.ds = Ctx.ds = Ctx.ss; \
382	\
383	/* \
384	* Loop over the test data and feed it to the worker. \
385	*/\
386	for (iTestData = 0; iTestData < cTestData; iTestData++) \
387	{ \
388	unsigned iRecompiler; \
389	a_UIntType const uSrc = !fMaskSrcWhenMemDst \| !paTests[iTest].fDstMem \
390	? paTestData[iTestData].uSrc2 : paTestData[iTestData].uSrc2 & (a_cBits - 1); \
391	if (!paTests[iTest].cBitsImm) \
392	{ \
393	*puCtxSrc = uSrc; \
394	*puCtxExpectSrc = uSrc; \
395	} \
396	else if (paTests[iTest].cBitsImm == 8) \
397	{ \
398	if ((int8_t)uSrc == (int##a_cBits##_t)uSrc) \
399	*pbImm = (uint8_t)uSrc; \
400	else continue; \
401	} \
402	else if (sizeof(puImm) == sizeof(puCtxSrc) \|\| (int32_t)uSrc == (int64_t)uSrc) \
403	*puImm = (a_UIntImmType)uSrc; \
404	else continue; \
405	\
406	*puCtxDst = paTestData[iTestData].uSrc1; \
407	*puCtxExpectDst = paTestData[iTestData].uResult; \
408	if (a_cBits == 32 && !fReadOnly && !paTests[iTest].fDstMem) \
409	puCtxExpectDst[1] = 0; \
410	\
411	if (puMemPtrReg) \
412	{ \
413	*puMemPtrReg = BS3_FP_OFF(&Buf.uData); \
414	*puMemPtrRegExpt = BS3_FP_OFF(&Buf.uData); \
415	} \
416	\
417	CtxExpect.rflags.u16 &= ~X86_EFL_STATUS_BITS; \
418	CtxExpect.rflags.u16 \|= paTestData[iTestData].fEflOut & X86_EFL_STATUS_BITS; \
419	\
420	/* \
421	* Do input the eight EFLAGS variations three times, so we're sure to trigger \
422	* native recompilation of the test worker code. \
423	*/ \
424	for (iRecompiler = 0; iRecompiler < 2; iRecompiler++) \
425	{ \
426	unsigned iEflVar = 0; \
427	for (iEflVar = 0; iEflVar < RT_ELEMENTS(g_auEflStatusBitsVars); iEflVar++) \
428	{ \
429	if (paTests[iTest].fDstMem) \
430	*puCtxDst = paTestData[iTestData].uSrc1; \
431	\
432	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS; \
433	if (!fCarryIn) \
434	Ctx.rflags.u16 \|= g_auEflStatusBitsVars[iEflVar]; \
435	else \
436	Ctx.rflags.u16 \|= (g_auEflStatusBitsVars[iEflVar] & ~X86_EFL_CF) \
437	\| (paTestData[iTestData].fEflOut >> BS3CPUINSTR2BIN_EFL_CARRY_IN_BIT) & X86_EFL_CF; \
438	if (fPassthruEfl) \
439	CtxExpect.rflags.u16 = (CtxExpect.rflags.u16 & ~fPassthruEfl) \| (Ctx.rflags.u16 & fPassthruEfl); \
440	\
441	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame); \
442	if (TrapFrame.bXcpt != X86_XCPT_UD) \
443	{ \
444	Bs3TestFailedF("Expected #UD got %#x", TrapFrame.bXcpt); \
445	Bs3TrapPrintFrame(&TrapFrame); \
446	} \
447	else if (Bs3TestCheckRegCtxEx(&TrapFrame.Ctx, &CtxExpect, 0 /cbPcAdjust/, 0 /cbSpAdjust/, \
448	0 /fExtraEfl/, "mode", (iTest << 8) \| (iTestData & 0xff))) \
449	{ \
450	if (!puMemPtrReg) \
451	continue; \
452	if (paTests[iTest].fDstMem && Buf.uData != uMemExpect) \
453	Bs3TestPrintf("Wrong memory result: %" a_szFmt ", expected %" a_szFmt "\n", Buf.uData, uMemExpect); \
454	else if (!paTests[iTest].fDstMem && Buf.uData != uSrc) \
455	Bs3TestPrintf("Memory input result modified: %" a_szFmt ", expected %" a_szFmt "\n", Buf.uData, uSrc); \
456	else \
457	continue; \
458	} \
459	/else { Bs3RegCtxPrint(&Ctx); Bs3TrapPrintFrame(&TrapFrame); }/ \
460	Bs3TestPrintf(#a_cBits ": iTest=%u iData=%u: uSrc1=%" a_szFmt "%s uSrc2=%" a_szFmt "%s %s-> %" a_szFmt "\n", \
461	iTest, iTestData, paTestData[iTestData].uSrc1, paTests[iTest].fDstMem ? " mem" : "", \
462	paTestData[iTestData].uSrc2, paTests[iTest].fSrcMem ? " mem" : "", \
463	!fCarryIn ? "" : Ctx.rflags.u16 & X86_EFL_CF ? "CF " : "NC ", \
464	paTestData[iTestData].uResult); \
465	Bs3RegCtxPrint(&Ctx); Bs3TrapPrintFrame(&TrapFrame); \
466	ASMHalt(); \
467	iRecompiler = ~0U - 1; \
468	break; \
469	} \
470	} \
471	} \
472	\
473	/* Restore modified context registers (except EFLAGS). */ \
474	CtxExpect.ds = Ctx.ds = SavedDs; \
475	(&CtxExpect.rax)[idxDstReg & 15].u = (&Ctx.rax)[idxDstReg & 15].u = SavedDst.u; \
476	(&CtxExpect.rax)[idxSrcReg & 15].u = (&Ctx.rax)[idxSrcReg & 15].u = SavedSrc.u; \
477	} \
478	\
479	return 0; \
480	}
481
482	BS3CPUINSTR2_COMMON_BINARY_U(8, uint8_t, uint8_t, "RX8")
483	BS3CPUINSTR2_COMMON_BINARY_U(16, uint16_t, uint16_t, "RX16")
484	BS3CPUINSTR2_COMMON_BINARY_U(32, uint32_t, uint32_t, "RX32")
485	#if ARCH_BITS == 64
486	BS3CPUINSTR2_COMMON_BINARY_U(64, uint64_t, uint32_t, "RX64")
487	#endif
488
489
490	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_and)(uint8_t bMode)
491	{
492	static const BS3CPUINSTR2CMNBINTEST s_aTests8[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8(and) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8(and) };
493	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16(and) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(and) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W(and) };
494	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_32(and) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(and) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW(and) };
495	#if ARCH_BITS == 64
496	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_64(and) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(and) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64DW(and) };
497	#endif
498	bs3CpuInstr2_CommonBinaryU8(bMode, s_aTests8, RT_ELEMENTS(s_aTests8), 0 /fPassthruEfl/,
499	g_aBs3CpuInstr2_and_TestDataU8, g_cBs3CpuInstr2_and_TestDataU8,
500	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
501	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), 0 /fPassthruEfl/,
502	g_aBs3CpuInstr2_and_TestDataU16, g_cBs3CpuInstr2_and_TestDataU16,
503	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
504	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), 0 /fPassthruEfl/,
505	g_aBs3CpuInstr2_and_TestDataU32, g_cBs3CpuInstr2_and_TestDataU32,
506	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
507	#if ARCH_BITS == 64
508	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), 0 /fPassthruEfl/,
509	g_aBs3CpuInstr2_and_TestDataU64, g_cBs3CpuInstr2_and_TestDataU64,
510	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
511	#endif
512	return 0;
513	}
514
515
516	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_or)(uint8_t bMode)
517	{
518	static const BS3CPUINSTR2CMNBINTEST s_aTests8[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8(or) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8(or) };
519	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16(or) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(or) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W(or) };
520	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_32(or) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(or) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW(or) };
521	#if ARCH_BITS == 64
522	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_64(or) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(or) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64DW(or) };
523	#endif
524	bs3CpuInstr2_CommonBinaryU8(bMode, s_aTests8, RT_ELEMENTS(s_aTests8), 0 /fPassthruEfl/,
525	g_aBs3CpuInstr2_or_TestDataU8, g_cBs3CpuInstr2_or_TestDataU8,
526	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
527	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), 0 /fPassthruEfl/,
528	g_aBs3CpuInstr2_or_TestDataU16, g_cBs3CpuInstr2_or_TestDataU16,
529	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
530	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), 0 /fPassthruEfl/,
531	g_aBs3CpuInstr2_or_TestDataU32, g_cBs3CpuInstr2_or_TestDataU32,
532	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
533	#if ARCH_BITS == 64
534	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), 0 /fPassthruEfl/,
535	g_aBs3CpuInstr2_or_TestDataU64, g_cBs3CpuInstr2_or_TestDataU64,
536	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
537	#endif
538	return 0;
539	}
540
541
542	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_xor)(uint8_t bMode)
543	{
544	static const BS3CPUINSTR2CMNBINTEST s_aTests8[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8(xor) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8(xor) };
545	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16(xor) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(xor) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W(xor) };
546	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_32(xor) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(xor) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW(xor) };
547	#if ARCH_BITS == 64
548	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_64(xor) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(xor) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64DW(xor) };
549	#endif
550	bs3CpuInstr2_CommonBinaryU8(bMode, s_aTests8, RT_ELEMENTS(s_aTests8), 0 /fPassthruEfl/,
551	g_aBs3CpuInstr2_xor_TestDataU8, g_cBs3CpuInstr2_xor_TestDataU8,
552	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
553	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), 0 /fPassthruEfl/,
554	g_aBs3CpuInstr2_xor_TestDataU16, g_cBs3CpuInstr2_xor_TestDataU16,
555	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
556	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), 0 /fPassthruEfl/,
557	g_aBs3CpuInstr2_xor_TestDataU32, g_cBs3CpuInstr2_xor_TestDataU32,
558	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
559	#if ARCH_BITS == 64
560	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), 0 /fPassthruEfl/,
561	g_aBs3CpuInstr2_xor_TestDataU64, g_cBs3CpuInstr2_xor_TestDataU64,
562	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
563	#endif
564	return 0;
565	}
566
567
568	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_test)(uint8_t bMode)
569	{
570	static const BS3CPUINSTR2CMNBINTEST s_aTests8[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_8(test) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8(test) };
571	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_16(test) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W(test) };
572	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_32(test) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW(test) };
573	#if ARCH_BITS == 64
574	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_64(test) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64DW(test) };
575	#endif
576	bs3CpuInstr2_CommonBinaryU8(bMode, s_aTests8, RT_ELEMENTS(s_aTests8), 0 /fPassthruEfl/,
577	g_aBs3CpuInstr2_test_TestDataU8, g_cBs3CpuInstr2_test_TestDataU8,
578	false /fCarryIn/, false /fMaskSrcWhenMemDst/, true /fReadOnly/);
579	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), 0 /fPassthruEfl/,
580	g_aBs3CpuInstr2_test_TestDataU16, g_cBs3CpuInstr2_test_TestDataU16,
581	false /fCarryIn/, false /fMaskSrcWhenMemDst/, true /fReadOnly/);
582	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), 0 /fPassthruEfl/,
583	g_aBs3CpuInstr2_test_TestDataU32, g_cBs3CpuInstr2_test_TestDataU32,
584	false /fCarryIn/, false /fMaskSrcWhenMemDst/, true /fReadOnly/);
585	#if ARCH_BITS == 64
586	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), 0 /fPassthruEfl/,
587	g_aBs3CpuInstr2_test_TestDataU64, g_cBs3CpuInstr2_test_TestDataU64,
588	false /fCarryIn/, false /fMaskSrcWhenMemDst/, true /fReadOnly/);
589	#endif
590	return 0;
591	}
592
593
594	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_add)(uint8_t bMode)
595	{
596	static const BS3CPUINSTR2CMNBINTEST s_aTests8[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8(add) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8(add) };
597	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16(add) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(add) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W(add) };
598	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_32(add) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(add) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW(add) };
599	#if ARCH_BITS == 64
600	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_64(add) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(add) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64DW(add) };
601	#endif
602	bs3CpuInstr2_CommonBinaryU8(bMode, s_aTests8, RT_ELEMENTS(s_aTests8), 0 /fPassthruEfl/,
603	g_aBs3CpuInstr2_add_TestDataU8, g_cBs3CpuInstr2_add_TestDataU8,
604	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
605	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), 0 /fPassthruEfl/,
606	g_aBs3CpuInstr2_add_TestDataU16, g_cBs3CpuInstr2_add_TestDataU16,
607	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
608	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), 0 /fPassthruEfl/,
609	g_aBs3CpuInstr2_add_TestDataU32, g_cBs3CpuInstr2_add_TestDataU32,
610	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
611	#if ARCH_BITS == 64
612	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), 0 /fPassthruEfl/,
613	g_aBs3CpuInstr2_add_TestDataU64, g_cBs3CpuInstr2_add_TestDataU64,
614	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
615	#endif
616	return 0;
617	}
618
619
620	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_adc)(uint8_t bMode)
621	{
622	static const BS3CPUINSTR2CMNBINTEST s_aTests8[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8(adc) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8(adc) };
623	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16(adc) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(adc) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W(adc) };
624	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_32(adc) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(adc) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW(adc) };
625	#if ARCH_BITS == 64
626	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_64(adc) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(adc) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64DW(adc) };
627	#endif
628	bs3CpuInstr2_CommonBinaryU8(bMode, s_aTests8, RT_ELEMENTS(s_aTests8), 0 /fPassthruEfl/,
629	g_aBs3CpuInstr2_adc_TestDataU8, g_cBs3CpuInstr2_adc_TestDataU8,
630	true /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
631	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), 0 /fPassthruEfl/,
632	g_aBs3CpuInstr2_adc_TestDataU16, g_cBs3CpuInstr2_adc_TestDataU16,
633	true /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
634	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), 0 /fPassthruEfl/,
635	g_aBs3CpuInstr2_adc_TestDataU32, g_cBs3CpuInstr2_adc_TestDataU32,
636	true /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
637	#if ARCH_BITS == 64
638	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), 0 /fPassthruEfl/,
639	g_aBs3CpuInstr2_adc_TestDataU64, g_cBs3CpuInstr2_adc_TestDataU64,
640	true /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
641	#endif
642	return 0;
643	}
644
645
646	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_sub)(uint8_t bMode)
647	{
648	static const BS3CPUINSTR2CMNBINTEST s_aTests8[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8(sub) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8(sub) };
649	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16(sub) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(sub) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W(sub) };
650	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_32(sub) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(sub) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW(sub) };
651	#if ARCH_BITS == 64
652	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_64(sub) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(sub) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64DW(sub) };
653	#endif
654	bs3CpuInstr2_CommonBinaryU8(bMode, s_aTests8, RT_ELEMENTS(s_aTests8), 0 /fPassthruEfl/,
655	g_aBs3CpuInstr2_sub_TestDataU8, g_cBs3CpuInstr2_sub_TestDataU8,
656	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
657	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), 0 /fPassthruEfl/,
658	g_aBs3CpuInstr2_sub_TestDataU16, g_cBs3CpuInstr2_sub_TestDataU16,
659	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
660	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), 0 /fPassthruEfl/,
661	g_aBs3CpuInstr2_sub_TestDataU32, g_cBs3CpuInstr2_sub_TestDataU32,
662	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
663	#if ARCH_BITS == 64
664	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), 0 /fPassthruEfl/,
665	g_aBs3CpuInstr2_sub_TestDataU64, g_cBs3CpuInstr2_sub_TestDataU64,
666	false /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
667	#endif
668	return 0;
669	}
670
671
672	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_sbb)(uint8_t bMode)
673	{
674	static const BS3CPUINSTR2CMNBINTEST s_aTests8[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8(sbb) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8(sbb) };
675	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16(sbb) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(sbb) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W(sbb) };
676	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_32(sbb) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(sbb) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW(sbb) };
677	#if ARCH_BITS == 64
678	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_64(sbb) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(sbb) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64DW(sbb) };
679	#endif
680	bs3CpuInstr2_CommonBinaryU8(bMode, s_aTests8, RT_ELEMENTS(s_aTests8), 0 /fPassthruEfl/,
681	g_aBs3CpuInstr2_sbb_TestDataU8, g_cBs3CpuInstr2_sbb_TestDataU8,
682	true /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
683	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), 0 /fPassthruEfl/,
684	g_aBs3CpuInstr2_sbb_TestDataU16, g_cBs3CpuInstr2_sbb_TestDataU16,
685	true /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
686	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), 0 /fPassthruEfl/,
687	g_aBs3CpuInstr2_sbb_TestDataU32, g_cBs3CpuInstr2_sbb_TestDataU32,
688	true /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
689	#if ARCH_BITS == 64
690	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), 0 /fPassthruEfl/,
691	g_aBs3CpuInstr2_sbb_TestDataU64, g_cBs3CpuInstr2_sbb_TestDataU64,
692	true /fCarryIn/, false /fMaskSrcWhenMemDst/, false /fReadOnly/);
693	#endif
694	return 0;
695	}
696
697
698	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_cmp)(uint8_t bMode)
699	{
700	static const BS3CPUINSTR2CMNBINTEST s_aTests8[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_8(cmp) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_8(cmp) };
701	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_16(cmp) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(cmp) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16W(cmp) };
702	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_32(cmp) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(cmp) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32DW(cmp) };
703	#if ARCH_BITS == 64
704	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_ALT_64(cmp) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(cmp) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64DW(cmp) };
705	#endif
706	bs3CpuInstr2_CommonBinaryU8(bMode, s_aTests8, RT_ELEMENTS(s_aTests8), 0 /fPassthruEfl/,
707	g_aBs3CpuInstr2_cmp_TestDataU8, g_cBs3CpuInstr2_cmp_TestDataU8,
708	false /fCarryIn/, false /fMaskSrcWhenMemDst/, true /fReadOnly/);
709	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), 0 /fPassthruEfl/,
710	g_aBs3CpuInstr2_cmp_TestDataU16, g_cBs3CpuInstr2_cmp_TestDataU16,
711	false /fCarryIn/, false /fMaskSrcWhenMemDst/, true /fReadOnly/);
712	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), 0 /fPassthruEfl/,
713	g_aBs3CpuInstr2_cmp_TestDataU32, g_cBs3CpuInstr2_cmp_TestDataU32,
714	false /fCarryIn/, false /fMaskSrcWhenMemDst/, true /fReadOnly/);
715	#if ARCH_BITS == 64
716	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), 0 /fPassthruEfl/,
717	g_aBs3CpuInstr2_cmp_TestDataU64, g_cBs3CpuInstr2_cmp_TestDataU64,
718	false /fCarryIn/, false /fMaskSrcWhenMemDst/, true /fReadOnly/);
719	#endif
720	return 0;
721	}
722
723
724	#define BS3CPUINSTR2_BTx_PASSTHRU_EFL (X86_EFL_ZF \| X86_EFL_PF \| X86_EFL_AF \| X86_EFL_SF \| X86_EFL_OF)
725
726	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_bt)(uint8_t bMode)
727	{
728	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_16(bt) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(bt) };
729	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_32(bt) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(bt) };
730	#if ARCH_BITS == 64
731	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_64(bt) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(bt) };
732	#endif
733	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
734	g_aBs3CpuInstr2_bt_TestDataU16, g_cBs3CpuInstr2_bt_TestDataU16,
735	false /fCarryIn/, true /fMaskSrcWhenMemDst/, true /fReadOnly/);
736	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
737	g_aBs3CpuInstr2_bt_TestDataU32, g_cBs3CpuInstr2_bt_TestDataU32,
738	false /fCarryIn/, true /fMaskSrcWhenMemDst/, true /fReadOnly/);
739	#if ARCH_BITS == 64
740	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
741	g_aBs3CpuInstr2_bt_TestDataU64, g_cBs3CpuInstr2_bt_TestDataU64,
742	false /fCarryIn/, true /fMaskSrcWhenMemDst/, true /fReadOnly/);
743	#endif
744	return 0;
745	}
746
747
748	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_btc)(uint8_t bMode)
749	{
750	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_16(btc) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(btc) };
751	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_32(btc) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(btc) };
752	#if ARCH_BITS == 64
753	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_64(btc) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(btc) };
754	#endif
755	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
756	g_aBs3CpuInstr2_btc_TestDataU16, g_cBs3CpuInstr2_btc_TestDataU16,
757	false /fCarryIn/, true /fMaskSrcWhenMemDst/, false /fReadOnly/);
758	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
759	g_aBs3CpuInstr2_btc_TestDataU32, g_cBs3CpuInstr2_btc_TestDataU32,
760	false /fCarryIn/, true /fMaskSrcWhenMemDst/, false /fReadOnly/);
761	#if ARCH_BITS == 64
762	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
763	g_aBs3CpuInstr2_btc_TestDataU64, g_cBs3CpuInstr2_btc_TestDataU64,
764	false /fCarryIn/, true /fMaskSrcWhenMemDst/, false /fReadOnly/);
765	#endif
766	return 0;
767	}
768
769
770	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_btr)(uint8_t bMode)
771	{
772	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_16(btr) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(btr) };
773	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_32(btr) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(btr) };
774	#if ARCH_BITS == 64
775	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_64(btr) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(btr) };
776	#endif
777	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
778	g_aBs3CpuInstr2_btr_TestDataU16, g_cBs3CpuInstr2_btr_TestDataU16,
779	false /fCarryIn/, true /fMaskSrcWhenMemDst/, false /fReadOnly/);
780	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
781	g_aBs3CpuInstr2_btr_TestDataU32, g_cBs3CpuInstr2_btr_TestDataU32,
782	false /fCarryIn/, true /fMaskSrcWhenMemDst/, false /fReadOnly/);
783	#if ARCH_BITS == 64
784	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
785	g_aBs3CpuInstr2_btr_TestDataU64, g_cBs3CpuInstr2_btr_TestDataU64,
786	false /fCarryIn/, true /fMaskSrcWhenMemDst/, false /fReadOnly/);
787	#endif
788	return 0;
789	}
790
791
792	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_bts)(uint8_t bMode)
793	{
794	static const BS3CPUINSTR2CMNBINTEST s_aTests16[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_16(bts) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_16B(bts) };
795	static const BS3CPUINSTR2CMNBINTEST s_aTests32[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_32(bts) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_32B(bts) };
796	#if ARCH_BITS == 64
797	static const BS3CPUINSTR2CMNBINTEST s_aTests64[] = { BS3CPUINSTR2CMNBINTEST_ENTRIES_64(bts) BS3CPUINSTR2CMNBINTEST_ENTRIES_IMM_64B(bts) };
798	#endif
799	bs3CpuInstr2_CommonBinaryU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
800	g_aBs3CpuInstr2_bts_TestDataU16, g_cBs3CpuInstr2_bts_TestDataU16,
801	false /fCarryIn/, true /fMaskSrcWhenMemDst/, false /fReadOnly/);
802	bs3CpuInstr2_CommonBinaryU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
803	g_aBs3CpuInstr2_bts_TestDataU32, g_cBs3CpuInstr2_bts_TestDataU32,
804	false /fCarryIn/, true /fMaskSrcWhenMemDst/, false /fReadOnly/);
805	#if ARCH_BITS == 64
806	bs3CpuInstr2_CommonBinaryU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), BS3CPUINSTR2_BTx_PASSTHRU_EFL,
807	g_aBs3CpuInstr2_bts_TestDataU64, g_cBs3CpuInstr2_bts_TestDataU64,
808	false /fCarryIn/, true /fMaskSrcWhenMemDst/, false /fReadOnly/);
809	#endif
810	return 0;
811	}
812
813
814
815	/*
816	* Basic shift & rotate tests.
817	*/
818
819	# if ARCH_BITS == 64 /* fDstMem cBitsImm */
820	# define BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_8_64BIT(a_Ins) \
821	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _sil_1), X86_GREG_xSI, 1, false }, \
822	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r9b_Ib), X86_GREG_x9, 8, false }, \
823	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r13b_cl), X86_GREG_x13, 0, false }, \
824	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _bDSx14_1), X86_GREG_x14, 1, true }, \
825	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _bDSxAX_Ib), X86_GREG_xAX, 8, true }, \
826	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _bDSx9_cl), X86_GREG_x9, 0, true },
827
828	# define BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_16_64BIT(a_Ins) \
829	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8w_1), X86_GREG_x8, 1, false }, \
830	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r9w_Ib), X86_GREG_x9, 8, false }, \
831	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r13w_cl), X86_GREG_x13, 0, false }, \
832	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _wDSx14_1), X86_GREG_x14, 1, true }, \
833	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _wDSxBP_Ib), X86_GREG_xBP, 8, true }, \
834	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _wDSx9_cl), X86_GREG_x9, 0, true },
835
836	# define BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_32_64BIT(a_Ins) \
837	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8d_1), X86_GREG_x8, 1, false }, \
838	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r9d_Ib), X86_GREG_x9, 8, false }, \
839	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r13d_cl), X86_GREG_x13, 0, false }, \
840	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dwDSx14_1), X86_GREG_x14, 1, true }, \
841	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dwDSxBP_Ib), X86_GREG_xBP, 8, true }, \
842	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dwDSx9_cl), X86_GREG_x9, 0, true },
843
844	# define BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_64(a_Ins) \
845	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _rdi_1), X86_GREG_xDI, 1, false }, \
846	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _rcx_Ib), X86_GREG_xCX, 8, false }, \
847	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _rbp_cl), X86_GREG_xBP, 0, false }, \
848	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _qwDSxSI_1), X86_GREG_xSI, 1, true }, \
849	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _qwDSxBX_Ib), X86_GREG_xBX, 8, true }, \
850	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _qwDSxDI_cl), X86_GREG_xDI, 0, true }, \
851	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r8_1), X86_GREG_x8, 1, false }, \
852	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r9_Ib), X86_GREG_x9, 8, false }, \
853	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _r13_cl), X86_GREG_x13, 0, false }, \
854	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _qwDSx14_1), X86_GREG_x14, 1, true }, \
855	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _qwDSxBP_Ib), X86_GREG_xBP, 8, true }, \
856	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _qwDSx9_cl), X86_GREG_x9, 0, true },
857
858	# else
859	# define BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_8_64BIT(aIns)
860	# define BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_16_64BIT(aIns)
861	# define BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_32_64BIT(aIns)
862	# endif
863
864	# define BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_8(a_Ins) \
865	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _al_1), X86_GREG_xAX, 1, false }, \
866	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dl_Ib), X86_GREG_xDX, 8, false }, \
867	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ch_cl), X86_GREG_xCX+16, 0, false }, \
868	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _bDSxBX_1), X86_GREG_xBX, 1, true }, \
869	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _bDSxDI_Ib), X86_GREG_xDI, 8, true }, \
870	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _bDSxSI_cl), X86_GREG_xSI, 0, true }, \
871	BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_8_64BIT(a_Ins)
872
873	# define BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_16(a_Ins) \
874	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _di_1), X86_GREG_xDI, 1, false }, \
875	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _cx_Ib), X86_GREG_xCX, 8, false }, \
876	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _bp_cl), X86_GREG_xBP, 0, false }, \
877	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _wDSxSI_1), X86_GREG_xSI, 1, true }, \
878	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _wDSxDI_Ib), X86_GREG_xDI, 8, true }, \
879	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _wDSxBX_cl), X86_GREG_xBX, 0, true }, \
880	BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_16_64BIT(a_Ins)
881
882	# define BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_32(a_Ins) \
883	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _edi_1), X86_GREG_xDI, 1, false }, \
884	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ecx_Ib), X86_GREG_xCX, 8, false }, \
885	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _ebp_cl), X86_GREG_xBP, 0, false }, \
886	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dwDSxSI_1), X86_GREG_xSI, 1, true }, \
887	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dwDSxBX_Ib), X86_GREG_xBX, 8, true }, \
888	{ BS3_CMN_NM(bs3CpuInstr2_ ## a_Ins ## _dwDSxDI_cl), X86_GREG_xDI, 0, true }, \
889	BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_32_64BIT(a_Ins)
890
891
892	typedef struct BS3CPUINSTR2CMNSHIFTTEST
893	{
894	FPFNBS3FAR pfnWorker;
895	uint8_t idxDstReg;
896	uint8_t cBitsImm; /*< 0 for CL; 1 for single shift w/o immediate; 8 for 8-bit immediate. /
897	bool fDstMem;
898	} BS3CPUINSTR2CMNSHIFTTEST;
899	typedef BS3CPUINSTR2CMNSHIFTTEST const BS3_FAR_DATA *PCBS3CPUINSTR2CMNSHIFTTEST;
900
901
902	DECLINLINE(uint16_t) bs3CpuInstr2_UndefEflByCpuVendor(BS3CPUVENDOR enmDataVendor, uint16_t fUndefEfl)
903	{
904	BS3CPUVENDOR enmVendor = Bs3GetCpuVendor();
905	return enmDataVendor == enmVendor
906	\|\| (enmDataVendor == BS3CPUVENDOR_INTEL) == (enmVendor != BS3CPUVENDOR_AMD && enmVendor != BS3CPUVENDOR_HYGON)
907	? 0 : fUndefEfl;
908	}
909
910
911	#define BS3CPUINSTR2_COMMON_SHIFT_U(a_cBits, a_UIntType, a_szFmt) \
912	static uint8_t \
913	RT_CONCAT(bs3CpuInstr2_CommonShiftU,a_cBits)(uint8_t bMode, PCBS3CPUINSTR2CMNSHIFTTEST paTests, unsigned cTests, \
914	RT_CONCAT(PCBS3CPUINSTR2SHIFT,a_cBits) paTestData, unsigned cTestData, \
915	uint16_t fUndefEfl, bool fIntelIbProblem) \
916	{ \
917	BS3REGCTX Ctx; \
918	BS3REGCTX CtxExpect; \
919	BS3TRAPFRAME TrapFrame; \
920	unsigned iTest; \
921	struct \
922	{ \
923	char achPreGuard[8]; \
924	a_UIntType uData; \
925	char achPostGuard[8]; \
926	} Buf = { { '0','1','2','3','4','5','6','7' }, 0, { '8','9','a','b','c','d','e','f'} }; \
927	a_UIntType uMemExpect = 0; \
928	uint8_t bMemDummy = 0; \
929	\
930	/* May have no test data for a CPU vendor*/ \
931	if (!cTestData) \
932	return 0; \
933	\
934	/* Ensure the structures are allocated before we sample the stack pointer. */ \
935	Bs3MemSet(&Ctx, 0, sizeof(Ctx)); \
936	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame)); \
937	\
938	/* \
939	* Create test context. \
940	*/ \
941	Bs3RegCtxSaveEx(&Ctx, bMode, 640); \
942	Ctx.rflags.u32 &= ~X86_EFL_RF; \
943	if (ARCH_BITS == 64) \
944	for (iTest = 0; iTest < 16; iTest++) \
945	if (iTest != X86_GREG_xSP) \
946	(&Ctx.rax)[iTest].au32[1] = UINT32_C(0x8572ade) << (iTest & 7); \
947	Bs3MemCpy(&CtxExpect, &Ctx, sizeof(CtxExpect)); \
948	if (!BS3_MODE_IS_16BIT_SYS(bMode)) \
949	CtxExpect.rflags.u32 \|= X86_EFL_RF; \
950	\
951	/* \
952	* Each test worker. \
953	*/ \
954	for (iTest = 0; iTest < cTests; iTest++) \
955	{ \
956	uint8_t const cbInstr = ((uint8_t BS3_FAR )paTests[iTest].pfnWorker)[-1]; / the function is prefixed by the length */ \
957	uint8_t RT_FAR * const pbImm = (uint8_t BS3_FAR )Code2RwPtr(&((uint8_t BS3_FAR )paTests[iTest].pfnWorker)[cbInstr - 1]); \
958	uint8_t const idxDstReg = paTests[iTest].idxDstReg; \
959	uint8_t const cBitsImm = paTests[iTest].cBitsImm; \
960	uint16_t const SavedDs = Ctx.ds; \
961	BS3REG const SavedDst = (&Ctx.rax)[idxDstReg & 15]; /* saves memptr too */ \
962	BS3REG const SavedRcx = Ctx.rcx; \
963	a_UIntType RT_FAR * const puCtxDst = paTests[iTest].fDstMem ? &Buf.uData \
964	: &(&Ctx.rax)[idxDstReg & 15].RT_CONCAT(au,a_cBits)[idxDstReg >> 4]; \
965	uint8_t RT_FAR * const puCtxSrc = cBitsImm == 0 ? &Ctx.rcx.au8[0] : &bMemDummy; \
966	a_UIntType RT_FAR * const puCtxExpectDst = paTests[iTest].fDstMem ? &uMemExpect \
967	: &(&CtxExpect.rax)[idxDstReg & 15].RT_CONCAT(au,a_cBits)[idxDstReg >> 4]; \
968	uint8_t RT_FAR * const puCtxExpectSrc = cBitsImm == 0 ? &CtxExpect.rcx.au8[0] : &bMemDummy; \
969	uint64_t RT_FAR * const puMemPtrReg = paTests[iTest].fDstMem ? &(&Ctx.rax)[idxDstReg & 15].u : NULL; \
970	uint64_t RT_FAR * const puMemPtrRegExpt= paTests[iTest].fDstMem ? &(&CtxExpect.rax)[idxDstReg & 15].u : NULL; \
971	unsigned cRecompOuter = 0; \
972	unsigned const cMaxRecompOuter= cBitsImm != 8 ? g_cBs3ThresholdNativeRecompiler + cTestData : 1; \
973	unsigned const cMaxRecompInner= cBitsImm != 8 ? 1 : g_cBs3ThresholdNativeRecompiler; \
974	/Bs3TestPrintf("\n"#a_cBits ": pfnWorker=%p cBitsImm=%d (%d)\n", paTests[iTest].pfnWorker, cBitsImm, paTests[iTest].cBitsImm);/ \
975	\
976	Bs3RegCtxSetRipCsFromLnkPtr(&Ctx, paTests[iTest].pfnWorker); \
977	CtxExpect.rip.u = Ctx.rip.u + cbInstr; \
978	CtxExpect.cs = Ctx.cs; \
979	\
980	if (puMemPtrReg) \
981	CtxExpect.ds = Ctx.ds = Ctx.ss; \
982	\
983	/* \
984	* Iterate twice or more over the input data to ensure that the recompiler kicks in. \
985	* For instructions with an immediate byte, we do this in the inner loop below. \
986	*/ \
987	while (cRecompOuter < cMaxRecompOuter) \
988	{ \
989	/* \
990	* Loop over the test data and feed it to the worker. \
991	*/\
992	unsigned iTestData; \
993	for (iTestData = 0; iTestData < cTestData; iTestData++) \
994	{ \
995	unsigned cRecompInner; \
996	uint8_t const uSrc2 = paTestData[iTestData].uSrc2; \
997	if (cBitsImm == 0) \
998	{ \
999	*puCtxSrc = uSrc2; \
1000	*puCtxExpectSrc = uSrc2; \
1001	} \
1002	else if (cBitsImm == 8) \
1003	*pbImm = uSrc2; \
1004	else if ((uSrc2 & RT_MAX(a_cBits - 1, 31)) != 1) \
1005	continue; \
1006	cRecompOuter++; \
1007	\
1008	*puCtxDst = paTestData[iTestData].uSrc1; \
1009	*puCtxExpectDst = paTestData[iTestData].uResult; \
1010	if (a_cBits == 32 && !paTests[iTest].fDstMem) \
1011	puCtxExpectDst[1] = 0; \
1012	\
1013	if (puMemPtrReg) \
1014	{ \
1015	*puMemPtrReg = BS3_FP_OFF(&Buf.uData); \
1016	*puMemPtrRegExpt = BS3_FP_OFF(&Buf.uData); \
1017	} \
1018	\
1019	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS; \
1020	Ctx.rflags.u16 \|= paTestData[iTestData].fEflIn & X86_EFL_STATUS_BITS; \
1021	CtxExpect.rflags.u16 &= ~X86_EFL_STATUS_BITS; \
1022	CtxExpect.rflags.u16 \|= paTestData[iTestData].fEflOut & X86_EFL_STATUS_BITS; \
1023	/* Intel: 'ROL reg,imm8' and 'ROR reg,imm8' produces different OF values. \
1024	stored in bit 3 of the output. Observed on 8700B, 9980HK, 10980xe, \
1025	1260p, ++. */ \
1026	if (fIntelIbProblem && cBitsImm == 8 && !paTests[iTest].fDstMem) \
1027	{ \
1028	CtxExpect.rflags.u16 &= ~X86_EFL_OF; \
1029	CtxExpect.rflags.u16 \|= (paTestData[iTestData].fEflOut & RT_BIT_32(BS3CPUINSTR2SHIFT_EFL_IB_OVERFLOW_OUT_BIT)) \
1030	<< (X86_EFL_OF_BIT - BS3CPUINSTR2SHIFT_EFL_IB_OVERFLOW_OUT_BIT); \
1031	} \
1032	\
1033	/* Inner recompiler trigger loop, for instructions with immediates that we modify. */ \
1034	cRecompInner = 0; \
1035	do \
1036	{ \
1037	if (paTests[iTest].fDstMem) \
1038	*puCtxDst = paTestData[iTestData].uSrc1; \
1039	\
1040	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame); \
1041	\
1042	if (fUndefEfl) /* When executing tests for the other CPU vendor. */ \
1043	CtxExpect.rflags.u16 = (CtxExpect.rflags.u16 & ~fUndefEfl) \| (TrapFrame.Ctx.rflags.u16 & fUndefEfl); \
1044	\
1045	if (TrapFrame.bXcpt != X86_XCPT_UD) \
1046	{ \
1047	Bs3TestFailedF("Expected #UD got %#x", TrapFrame.bXcpt); \
1048	Bs3TrapPrintFrame(&TrapFrame); \
1049	} \
1050	else if (Bs3TestCheckRegCtxEx(&TrapFrame.Ctx, &CtxExpect, 0 /cbPcAdjust/, 0 /cbSpAdjust/, \
1051	0 /fExtraEfl/, "mode", (iTest << 8) \| (iTestData & 0xff))) \
1052	{ \
1053	if (puMemPtrReg && paTests[iTest].fDstMem && Buf.uData != uMemExpect) \
1054	Bs3TestPrintf("Wrong memory result: %" a_szFmt ", expected %" a_szFmt "\n", Buf.uData, uMemExpect); \
1055	else \
1056	{ \
1057	cRecompInner++; \
1058	continue; \
1059	} \
1060	} \
1061	/else { Bs3RegCtxPrint(&Ctx); Bs3TrapPrintFrame(&TrapFrame); }/ \
1062	Bs3TestPrintf(#a_cBits ": iTest=%u iData=%u inner=%u: uSrc1=%" a_szFmt "%s uSrc2=%RX8 (%s) fEfl=%RX16 -> %" a_szFmt " fEfl=%RX16\n", \
1063	iTest, iTestData, cRecompInner, \
1064	paTestData[iTestData].uSrc1, paTests[iTest].fDstMem ? " mem" : "", \
1065	paTestData[iTestData].uSrc2, cBitsImm == 0 ? "cl" : cBitsImm == 1 ? "1" : "Ib", \
1066	(uint16_t)(paTestData[iTestData].fEflIn & X86_EFL_STATUS_BITS), \
1067	paTestData[iTestData].uResult, paTestData[iTestData].fEflOut & X86_EFL_STATUS_BITS); \
1068	Bs3RegCtxPrint(&Ctx); Bs3TrapPrintFrame(&TrapFrame); \
1069	ASMHalt(); \
1070	} while (cRecompInner < cMaxRecompInner); \
1071	} \
1072	} \
1073	\
1074	/* Restore modified context registers (except EFLAGS). */ \
1075	CtxExpect.ds = Ctx.ds = SavedDs; \
1076	(&CtxExpect.rax)[idxDstReg & 15].u = (&Ctx.rax)[idxDstReg & 15].u = SavedDst.u; \
1077	CtxExpect.rcx.u = Ctx.rcx.u = SavedRcx.u; \
1078	} \
1079	\
1080	return 0; \
1081	}
1082
1083	BS3CPUINSTR2_COMMON_SHIFT_U(8, uint8_t, "RX8")
1084	BS3CPUINSTR2_COMMON_SHIFT_U(16, uint16_t, "RX16")
1085	BS3CPUINSTR2_COMMON_SHIFT_U(32, uint32_t, "RX32")
1086	#if ARCH_BITS == 64
1087	BS3CPUINSTR2_COMMON_SHIFT_U(64, uint64_t, "RX64")
1088	#endif
1089
1090
1091	#define BS3CPUINSTR2_SHIFT_INSTR_NOT_64(a_Ins, a_fEflUndef, a_fIntelIbProblem) \
1092	{ \
1093	static const BS3CPUINSTR2CMNSHIFTTEST s_aTests8[] = { BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_8(a_Ins) }; \
1094	static const BS3CPUINSTR2CMNSHIFTTEST s_aTests16[] = { BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_16(a_Ins) }; \
1095	static const BS3CPUINSTR2CMNSHIFTTEST s_aTests32[] = { BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_32(a_Ins) }; \
1096	uint16_t const fEflUndefIntel = bs3CpuInstr2_UndefEflByCpuVendor(BS3CPUVENDOR_INTEL, a_fEflUndef); \
1097	uint16_t const fEflUndefAmd = bs3CpuInstr2_UndefEflByCpuVendor(BS3CPUVENDOR_AMD, a_fEflUndef); \
1098	bs3CpuInstr2_CommonShiftU8(bMode, s_aTests8, RT_ELEMENTS(s_aTests8), \
1099	g_aBs3CpuInstr2_ ## a_Ins ## _intel_TestDataU8, \
1100	g_cBs3CpuInstr2_ ## a_Ins ## _intel_TestDataU8, \
1101	fEflUndefIntel, a_fIntelIbProblem); \
1102	bs3CpuInstr2_CommonShiftU8(bMode, s_aTests8, RT_ELEMENTS(s_aTests8), \
1103	g_aBs3CpuInstr2_ ## a_Ins ## _amd_TestDataU8, \
1104	g_cBs3CpuInstr2_ ## a_Ins ## _amd_TestDataU8, \
1105	fEflUndefAmd, false); \
1106	bs3CpuInstr2_CommonShiftU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), \
1107	g_aBs3CpuInstr2_ ## a_Ins ## _intel_TestDataU16, \
1108	g_cBs3CpuInstr2_ ## a_Ins ## _intel_TestDataU16, \
1109	fEflUndefIntel, a_fIntelIbProblem); \
1110	bs3CpuInstr2_CommonShiftU16(bMode, s_aTests16, RT_ELEMENTS(s_aTests16), \
1111	g_aBs3CpuInstr2_ ## a_Ins ## _amd_TestDataU16, \
1112	g_cBs3CpuInstr2_ ## a_Ins ## _amd_TestDataU16, \
1113	fEflUndefAmd, false); \
1114	bs3CpuInstr2_CommonShiftU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), \
1115	g_aBs3CpuInstr2_ ## a_Ins ## _intel_TestDataU32, \
1116	g_cBs3CpuInstr2_ ## a_Ins ## _intel_TestDataU32, \
1117	fEflUndefIntel, a_fIntelIbProblem); \
1118	bs3CpuInstr2_CommonShiftU32(bMode, s_aTests32, RT_ELEMENTS(s_aTests32), \
1119	g_aBs3CpuInstr2_ ## a_Ins ## _amd_TestDataU32, \
1120	g_cBs3CpuInstr2_ ## a_Ins ## _amd_TestDataU32, \
1121	fEflUndefAmd, false); \
1122	} (void)0
1123	#if ARCH_BITS == 64
1124	# define BS3CPUINSTR2_SHIFT_INSTR_ONLY64(a_Ins, a_fEflUndef, a_fIntelIbProblem) \
1125	{ \
1126	static const BS3CPUINSTR2CMNSHIFTTEST s_aTests64[] = { BS3CPUINSTR2CMNSHIFTTEST_ENTRIES_64(a_Ins) }; \
1127	bs3CpuInstr2_CommonShiftU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), \
1128	g_aBs3CpuInstr2_ ## a_Ins ## _intel_TestDataU64, \
1129	g_cBs3CpuInstr2_ ## a_Ins ## _intel_TestDataU64, \
1130	bs3CpuInstr2_UndefEflByCpuVendor(BS3CPUVENDOR_INTEL, a_fEflUndef), a_fIntelIbProblem); \
1131	bs3CpuInstr2_CommonShiftU64(bMode, s_aTests64, RT_ELEMENTS(s_aTests64), \
1132	g_aBs3CpuInstr2_ ## a_Ins ## _amd_TestDataU64, \
1133	g_cBs3CpuInstr2_ ## a_Ins ## _amd_TestDataU64, \
1134	bs3CpuInstr2_UndefEflByCpuVendor(BS3CPUVENDOR_AMD, a_fEflUndef), false); \
1135	} (void)0
1136	#else
1137	# define BS3CPUINSTR2_SHIFT_INSTR_ONLY64(a_Ins, a_fEflUndef, a_fIntelIbProblem) (void)0
1138	#endif
1139
1140
1141	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_shl)(uint8_t bMode)
1142	{
1143	BS3CPUINSTR2_SHIFT_INSTR_NOT_64(shl, X86_EFL_AF \| X86_EFL_OF, false);
1144	BS3CPUINSTR2_SHIFT_INSTR_ONLY64(shl, X86_EFL_AF \| X86_EFL_OF, false);
1145	return 0;
1146	}
1147
1148
1149	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_shr)(uint8_t bMode)
1150	{
1151	BS3CPUINSTR2_SHIFT_INSTR_NOT_64(shr, X86_EFL_AF \| X86_EFL_OF, false);
1152	BS3CPUINSTR2_SHIFT_INSTR_ONLY64(shr, X86_EFL_AF \| X86_EFL_OF, false);
1153	return 0;
1154	}
1155
1156
1157	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_sar)(uint8_t bMode)
1158	{
1159	BS3CPUINSTR2_SHIFT_INSTR_NOT_64(sar, X86_EFL_AF \| X86_EFL_OF, false);
1160	BS3CPUINSTR2_SHIFT_INSTR_ONLY64(sar, X86_EFL_AF \| X86_EFL_OF, false);
1161	return 0;
1162	}
1163
1164
1165	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_rol)(uint8_t bMode)
1166	{
1167	BS3CPUINSTR2_SHIFT_INSTR_NOT_64(rol, X86_EFL_OF \| X86_EFL_CF, true /fIntelIbProblem/);
1168	BS3CPUINSTR2_SHIFT_INSTR_ONLY64(rol, X86_EFL_OF \| X86_EFL_CF, true /fIntelIbProblem/);
1169	return 0;
1170	}
1171
1172
1173	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_ror)(uint8_t bMode)
1174	{
1175	BS3CPUINSTR2_SHIFT_INSTR_NOT_64(ror, X86_EFL_OF \| X86_EFL_CF, true /fIntelIbProblem/);
1176	BS3CPUINSTR2_SHIFT_INSTR_ONLY64(ror, X86_EFL_OF \| X86_EFL_CF, true /fIntelIbProblem/);
1177	return 0;
1178	}
1179
1180
1181	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_rcl)(uint8_t bMode)
1182	{
1183	BS3CPUINSTR2_SHIFT_INSTR_NOT_64(rcl, X86_EFL_OF, false);
1184	BS3CPUINSTR2_SHIFT_INSTR_ONLY64(rcl, X86_EFL_OF, false);
1185	return 0;
1186	}
1187
1188
1189	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_rcr)(uint8_t bMode)
1190	{
1191	BS3CPUINSTR2_SHIFT_INSTR_NOT_64(rcr, X86_EFL_OF, false);
1192	BS3CPUINSTR2_SHIFT_INSTR_ONLY64(rcr, X86_EFL_OF, false);
1193	return 0;
1194	}
1195
1196
1197
1198
1199
1200	/*
1201	* Multiplication
1202	*/
1203
1204	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_mul)(uint8_t bMode)
1205	{
1206	#define MUL_CHECK_EFLAGS_ZERO (uint16_t)(X86_EFL_AF \| X86_EFL_ZF)
1207	#define MUL_CHECK_EFLAGS (uint16_t)(X86_EFL_CF \| X86_EFL_OF \| X86_EFL_SF \| X86_EFL_PF)
1208
1209	static const struct
1210	{
1211	RTCCUINTREG uInAX;
1212	RTCCUINTREG uInBX;
1213	RTCCUINTREG uOutDX;
1214	RTCCUINTREG uOutAX;
1215	uint16_t fFlags;
1216	} s_aTests[] =
1217	{
1218	{ 1, 1,
1219	0, 1, 0 },
1220	{ 2, 2,
1221	0, 4, 0 },
1222	{ RTCCUINTREG_MAX, RTCCUINTREG_MAX,
1223	RTCCUINTREG_MAX-1, 1, X86_EFL_CF \| X86_EFL_OF },
1224	{ RTCCINTREG_MAX, RTCCINTREG_MAX,
1225	RTCCINTREG_MAX / 2, 1, X86_EFL_CF \| X86_EFL_OF },
1226	{ 1, RTCCUINTREG_MAX,
1227	0, RTCCUINTREG_MAX, X86_EFL_PF \| X86_EFL_SF },
1228	{ 1, RTCCINTREG_MAX,
1229	0, RTCCINTREG_MAX, X86_EFL_PF },
1230	{ 2, RTCCINTREG_MAX,
1231	0, RTCCUINTREG_MAX - 1, X86_EFL_SF },
1232	{ (RTCCUINTREG)RTCCINTREG_MAX + 1, 2,
1233	1, 0, X86_EFL_PF \| X86_EFL_CF \| X86_EFL_OF },
1234	{ (RTCCUINTREG)RTCCINTREG_MAX / 2 + 1, 3,
1235	0, ((RTCCUINTREG)RTCCINTREG_MAX / 2 + 1) * 3, X86_EFL_PF \| X86_EFL_SF },
1236	};
1237
1238	BS3REGCTX Ctx;
1239	BS3TRAPFRAME TrapFrame;
1240	unsigned i, j, k;
1241
1242	/* Ensure the structures are allocated before we sample the stack pointer. */
1243	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
1244	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
1245
1246	/*
1247	* Create test context.
1248	*/
1249	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
1250	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, BS3_CMN_NM(bs3CpuInstr2_mul_xBX_ud2));
1251	for (k = 0; k < 2; k++)
1252	{
1253	Ctx.rflags.u16 \|= MUL_CHECK_EFLAGS \| MUL_CHECK_EFLAGS_ZERO;
1254	for (j = 0; j < 2; j++)
1255	{
1256	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
1257	{
1258	if (k == 0)
1259	{
1260	Ctx.rax.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInAX;
1261	Ctx.rbx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInBX;
1262	}
1263	else
1264	{
1265	Ctx.rax.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInBX;
1266	Ctx.rbx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInAX;
1267	}
1268	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
1269	if (TrapFrame.bXcpt != X86_XCPT_UD)
1270	Bs3TestFailedF("Expected #UD got %#x", TrapFrame.bXcpt);
1271	else if ( TrapFrame.Ctx.rax.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutAX
1272	\|\| TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutDX
1273	\|\| (TrapFrame.Ctx.rflags.u16 & (MUL_CHECK_EFLAGS \| MUL_CHECK_EFLAGS_ZERO))
1274	!= (s_aTests[i].fFlags & MUL_CHECK_EFLAGS) )
1275	{
1276	Bs3TestFailedF("test #%i failed: input %#" RTCCUINTREG_XFMT " * %#" RTCCUINTREG_XFMT,
1277	i, s_aTests[i].uInAX, s_aTests[i].uInBX);
1278
1279	if (TrapFrame.Ctx.rax.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutAX)
1280	Bs3TestFailedF("Expected xAX = %#RX" RT_XSTR(ARCH_BITS) " got %#RX" RT_XSTR(ARCH_BITS),
1281	s_aTests[i].uOutAX, TrapFrame.Ctx.rax.RT_CONCAT(u,ARCH_BITS));
1282	if (TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutDX)
1283	Bs3TestFailedF("Expected xDX = %#RX" RT_XSTR(ARCH_BITS) " got %#RX" RT_XSTR(ARCH_BITS),
1284	s_aTests[i].uOutDX, TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS));
1285	if ( (TrapFrame.Ctx.rflags.u16 & (MUL_CHECK_EFLAGS \| MUL_CHECK_EFLAGS_ZERO))
1286	!= (s_aTests[i].fFlags & MUL_CHECK_EFLAGS) )
1287	Bs3TestFailedF("Expected EFLAGS = %#06RX16, got %#06RX16", s_aTests[i].fFlags & MUL_CHECK_EFLAGS,
1288	TrapFrame.Ctx.rflags.u16 & (MUL_CHECK_EFLAGS \| MUL_CHECK_EFLAGS_ZERO));
1289	}
1290	}
1291	Ctx.rflags.u16 &= ~(MUL_CHECK_EFLAGS \| MUL_CHECK_EFLAGS_ZERO);
1292	}
1293	}
1294
1295	return 0;
1296	}
1297
1298
1299	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_imul)(uint8_t bMode)
1300	{
1301	#define IMUL_CHECK_EFLAGS_ZERO (uint16_t)(X86_EFL_AF \| X86_EFL_ZF)
1302	#define IMUL_CHECK_EFLAGS (uint16_t)(X86_EFL_CF \| X86_EFL_OF \| X86_EFL_SF \| X86_EFL_PF)
1303	static const struct
1304	{
1305	RTCCUINTREG uInAX;
1306	RTCCUINTREG uInBX;
1307	RTCCUINTREG uOutDX;
1308	RTCCUINTREG uOutAX;
1309	uint16_t fFlags;
1310	} s_aTests[] =
1311	{
1312	/* two positive values. */
1313	{ 1, 1,
1314	0, 1, 0 },
1315	{ 2, 2,
1316	0, 4, 0 },
1317	{ RTCCINTREG_MAX, RTCCINTREG_MAX,
1318	RTCCINTREG_MAX/2, 1, X86_EFL_CF \| X86_EFL_OF },
1319	{ 1, RTCCINTREG_MAX,
1320	0, RTCCINTREG_MAX, X86_EFL_PF },
1321	{ 2, RTCCINTREG_MAX,
1322	0, RTCCUINTREG_MAX - 1U, X86_EFL_CF \| X86_EFL_OF \| X86_EFL_SF },
1323	{ 2, RTCCINTREG_MAX / 2,
1324	0, RTCCINTREG_MAX - 1U, 0 },
1325	{ 2, (RTCCINTREG_MAX / 2 + 1),
1326	0, (RTCCUINTREG)RTCCINTREG_MAX + 1U, X86_EFL_CF \| X86_EFL_OF \| X86_EFL_SF \| X86_EFL_PF },
1327	{ 4, (RTCCINTREG_MAX / 2 + 1),
1328	1, 0, X86_EFL_CF \| X86_EFL_OF \| X86_EFL_PF },
1329
1330	/* negative and positive */
1331	{ -4, 3,
1332	-1, -12, X86_EFL_SF },
1333	{ 32, -127,
1334	-1, -4064, X86_EFL_SF },
1335	{ RTCCINTREG_MIN, 1,
1336	-1, RTCCINTREG_MIN, X86_EFL_SF \| X86_EFL_PF },
1337	{ RTCCINTREG_MIN, 2,
1338	-1, 0, X86_EFL_CF \| X86_EFL_OF \| X86_EFL_PF },
1339	{ RTCCINTREG_MIN, 3,
1340	-2, RTCCINTREG_MIN, X86_EFL_CF \| X86_EFL_OF \| X86_EFL_SF \| X86_EFL_PF },
1341	{ RTCCINTREG_MIN, 4,
1342	-2, 0, X86_EFL_CF \| X86_EFL_OF \| X86_EFL_PF },
1343	{ RTCCINTREG_MIN, RTCCINTREG_MAX,
1344	RTCCINTREG_MIN / 2, RTCCINTREG_MIN, X86_EFL_CF \| X86_EFL_OF \| X86_EFL_SF \| X86_EFL_PF },
1345	{ RTCCINTREG_MIN, RTCCINTREG_MAX - 1,
1346	RTCCINTREG_MIN / 2 + 1, 0, X86_EFL_CF \| X86_EFL_OF \| X86_EFL_PF },
1347
1348	/* two negative values. */
1349	{ -4, -63,
1350	0, 252, X86_EFL_PF },
1351	{ RTCCINTREG_MIN, RTCCINTREG_MIN,
1352	RTCCUINTREG_MAX / 4 + 1, 0, X86_EFL_CF \| X86_EFL_OF \| X86_EFL_PF },
1353	{ RTCCINTREG_MIN, RTCCINTREG_MIN + 1,
1354	RTCCUINTREG_MAX / 4, RTCCINTREG_MIN, X86_EFL_CF \| X86_EFL_OF \| X86_EFL_SF \| X86_EFL_PF},
1355	{ RTCCINTREG_MIN + 1, RTCCINTREG_MIN + 1,
1356	RTCCUINTREG_MAX / 4, 1, X86_EFL_CF \| X86_EFL_OF },
1357
1358	};
1359
1360	BS3REGCTX Ctx;
1361	BS3TRAPFRAME TrapFrame;
1362	unsigned i, j, k;
1363
1364	/* Ensure the structures are allocated before we sample the stack pointer. */
1365	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
1366	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
1367
1368	/*
1369	* Create test context.
1370	*/
1371	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
1372	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, BS3_CMN_NM(bs3CpuInstr2_imul_xBX_ud2));
1373
1374	for (k = 0; k < 2; k++)
1375	{
1376	Ctx.rflags.u16 \|= MUL_CHECK_EFLAGS \| MUL_CHECK_EFLAGS_ZERO;
1377	for (j = 0; j < 2; j++)
1378	{
1379	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
1380	{
1381	if (k == 0)
1382	{
1383	Ctx.rax.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInAX;
1384	Ctx.rbx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInBX;
1385	}
1386	else
1387	{
1388	Ctx.rax.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInBX;
1389	Ctx.rbx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInAX;
1390	}
1391	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
1392	if (TrapFrame.bXcpt != X86_XCPT_UD)
1393	Bs3TestFailedF("Expected #UD got %#x", TrapFrame.bXcpt);
1394	else if ( TrapFrame.Ctx.rax.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutAX
1395	\|\| TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutDX
1396	\|\| (TrapFrame.Ctx.rflags.u16 & (IMUL_CHECK_EFLAGS \| IMUL_CHECK_EFLAGS_ZERO))
1397	!= (s_aTests[i].fFlags & IMUL_CHECK_EFLAGS) )
1398	{
1399	Bs3TestFailedF("test #%i failed: input %#" RTCCUINTREG_XFMT " * %#" RTCCUINTREG_XFMT,
1400	i, s_aTests[i].uInAX, s_aTests[i].uInBX);
1401
1402	if (TrapFrame.Ctx.rax.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutAX)
1403	Bs3TestFailedF("Expected xAX = %#RX" RT_XSTR(ARCH_BITS) " got %#RX" RT_XSTR(ARCH_BITS),
1404	s_aTests[i].uOutAX, TrapFrame.Ctx.rax.RT_CONCAT(u,ARCH_BITS));
1405	if (TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutDX)
1406	Bs3TestFailedF("Expected xDX = %#RX" RT_XSTR(ARCH_BITS) " got %#RX" RT_XSTR(ARCH_BITS),
1407	s_aTests[i].uOutDX, TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS));
1408	if ( (TrapFrame.Ctx.rflags.u16 & (IMUL_CHECK_EFLAGS \| IMUL_CHECK_EFLAGS_ZERO))
1409	!= (s_aTests[i].fFlags & IMUL_CHECK_EFLAGS) )
1410	Bs3TestFailedF("Expected EFLAGS = %#06RX16, got %#06RX16", s_aTests[i].fFlags & IMUL_CHECK_EFLAGS,
1411	TrapFrame.Ctx.rflags.u16 & (IMUL_CHECK_EFLAGS \| IMUL_CHECK_EFLAGS_ZERO));
1412	}
1413	}
1414	}
1415	}
1416
1417	/*
1418	* Repeat for the truncating two operand version.
1419	*/
1420	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, BS3_CMN_NM(bs3CpuInstr2_imul_xCX_xBX_ud2));
1421
1422	for (k = 0; k < 2; k++)
1423	{
1424	Ctx.rflags.u16 \|= MUL_CHECK_EFLAGS \| MUL_CHECK_EFLAGS_ZERO;
1425	for (j = 0; j < 2; j++)
1426	{
1427	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
1428	{
1429	if (k == 0)
1430	{
1431	Ctx.rcx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInAX;
1432	Ctx.rbx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInBX;
1433	}
1434	else
1435	{
1436	Ctx.rcx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInBX;
1437	Ctx.rbx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInAX;
1438	}
1439	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
1440	if (TrapFrame.bXcpt != X86_XCPT_UD)
1441	Bs3TestFailedF("Expected #UD got %#x", TrapFrame.bXcpt);
1442	else if ( TrapFrame.Ctx.rcx.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutAX
1443	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
1444	\|\| TrapFrame.Ctx.rbx.u != Ctx.rbx.u
1445	\|\| (TrapFrame.Ctx.rflags.u16 & (IMUL_CHECK_EFLAGS \| IMUL_CHECK_EFLAGS_ZERO))
1446	!= (s_aTests[i].fFlags & IMUL_CHECK_EFLAGS) )
1447	{
1448	Bs3TestFailedF("test #%i failed: input %#" RTCCUINTREG_XFMT " * %#" RTCCUINTREG_XFMT,
1449	i, s_aTests[i].uInAX, s_aTests[i].uInBX);
1450
1451	if (TrapFrame.Ctx.rcx.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutAX)
1452	Bs3TestFailedF("Expected xAX = %#RX" RT_XSTR(ARCH_BITS) " got %#RX" RT_XSTR(ARCH_BITS),
1453	s_aTests[i].uOutAX, TrapFrame.Ctx.rcx.RT_CONCAT(u,ARCH_BITS));
1454	if ( (TrapFrame.Ctx.rflags.u16 & (IMUL_CHECK_EFLAGS \| IMUL_CHECK_EFLAGS_ZERO))
1455	!= (s_aTests[i].fFlags & IMUL_CHECK_EFLAGS) )
1456	Bs3TestFailedF("Expected EFLAGS = %#06RX16, got %#06RX16", s_aTests[i].fFlags & IMUL_CHECK_EFLAGS,
1457	TrapFrame.Ctx.rflags.u16 & (IMUL_CHECK_EFLAGS \| IMUL_CHECK_EFLAGS_ZERO));
1458	}
1459	}
1460	}
1461	}
1462
1463	return 0;
1464	}
1465
1466
1467	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_div)(uint8_t bMode)
1468	{
1469	#define DIV_CHECK_EFLAGS (uint16_t)(X86_EFL_CF \| X86_EFL_OF \| X86_EFL_SF \| X86_EFL_ZF \| X86_EFL_AF \| X86_EFL_PF)
1470	static const struct
1471	{
1472	RTCCUINTREG uInDX;
1473	RTCCUINTREG uInAX;
1474	RTCCUINTREG uInBX;
1475	RTCCUINTREG uOutAX;
1476	RTCCUINTREG uOutDX;
1477	uint8_t bXcpt;
1478	} s_aTests[] =
1479	{
1480	{ 0, 1, 1,
1481	1, 0, X86_XCPT_UD },
1482	{ 0, 5, 2,
1483	2, 1, X86_XCPT_UD },
1484	{ 0, 0, 0,
1485	0, 0, X86_XCPT_DE },
1486	{ RTCCUINTREG_MAX, RTCCUINTREG_MAX, 0,
1487	0, 0, X86_XCPT_DE },
1488	{ RTCCUINTREG_MAX, RTCCUINTREG_MAX, 1,
1489	0, 0, X86_XCPT_DE },
1490	{ RTCCUINTREG_MAX, RTCCUINTREG_MAX, RTCCUINTREG_MAX,
1491	0, 0, X86_XCPT_DE },
1492	{ RTCCUINTREG_MAX - 1, RTCCUINTREG_MAX, RTCCUINTREG_MAX,
1493	RTCCUINTREG_MAX, RTCCUINTREG_MAX - 1, X86_XCPT_UD },
1494	};
1495
1496	BS3REGCTX Ctx;
1497	BS3TRAPFRAME TrapFrame;
1498	unsigned i, j;
1499
1500	/* Ensure the structures are allocated before we sample the stack pointer. */
1501	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
1502	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
1503
1504	/*
1505	* Create test context.
1506	*/
1507	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
1508	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, BS3_CMN_NM(bs3CpuInstr2_div_xBX_ud2));
1509
1510	/*
1511	* Do the tests twice, first with all flags set, then once again with
1512	* flags cleared. The flags are not touched by my intel skylake CPU.
1513	*/
1514	Ctx.rflags.u16 \|= DIV_CHECK_EFLAGS;
1515	for (j = 0; j < 2; j++)
1516	{
1517	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
1518	{
1519	Ctx.rax.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInAX;
1520	Ctx.rdx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInDX;
1521	Ctx.rbx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInBX;
1522	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
1523
1524	if ( TrapFrame.bXcpt != s_aTests[i].bXcpt
1525	\|\| ( s_aTests[i].bXcpt == X86_XCPT_UD
1526	? TrapFrame.Ctx.rax.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutAX
1527	\|\| TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutDX
1528	\|\| (TrapFrame.Ctx.rflags.u16 & DIV_CHECK_EFLAGS) != (Ctx.rflags.u16 & DIV_CHECK_EFLAGS)
1529	: TrapFrame.Ctx.rax.u != Ctx.rax.u
1530	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
1531	\|\| (TrapFrame.Ctx.rflags.u16 & DIV_CHECK_EFLAGS) != (Ctx.rflags.u16 & DIV_CHECK_EFLAGS) ) )
1532	{
1533	Bs3TestFailedF("test #%i failed: input %#" RTCCUINTREG_XFMT ":%" RTCCUINTREG_XFMT " / %#" RTCCUINTREG_XFMT,
1534	i, s_aTests[i].uInDX, s_aTests[i].uInAX, s_aTests[i].uInBX);
1535	if (TrapFrame.bXcpt != s_aTests[i].bXcpt)
1536	Bs3TestFailedF("Expected bXcpt = %#x, got %#x", s_aTests[i].bXcpt, TrapFrame.bXcpt);
1537	if (s_aTests[i].bXcpt == X86_XCPT_UD)
1538	{
1539	if (TrapFrame.Ctx.rax.RT_CONCAT(u, ARCH_BITS) != s_aTests[i].uOutAX)
1540	Bs3TestFailedF("Expected xAX = %#" RTCCUINTREG_XFMT ", got %#" RTCCUINTREG_XFMT,
1541	s_aTests[i].uOutAX, TrapFrame.Ctx.rax.RT_CONCAT(u,ARCH_BITS));
1542	if (TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutDX)
1543	Bs3TestFailedF("Expected xDX = %#" RTCCUINTREG_XFMT ", got %#" RTCCUINTREG_XFMT,
1544	s_aTests[i].uOutDX, TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS));
1545	if ((TrapFrame.Ctx.rflags.u16 & DIV_CHECK_EFLAGS) != (Ctx.rflags.u16 & DIV_CHECK_EFLAGS))
1546	Bs3TestFailedF("Expected EFLAGS = %#06RX16, got %#06RX16",
1547	Ctx.rflags.u16 & DIV_CHECK_EFLAGS, TrapFrame.Ctx.rflags.u16 & DIV_CHECK_EFLAGS);
1548	}
1549	}
1550	}
1551	Ctx.rflags.u16 &= ~DIV_CHECK_EFLAGS;
1552	}
1553
1554	return 0;
1555	}
1556
1557
1558
1559	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_idiv)(uint8_t bMode)
1560	{
1561	#define IDIV_CHECK_EFLAGS (uint16_t)(X86_EFL_CF \| X86_EFL_OF \| X86_EFL_SF \| X86_EFL_ZF \| X86_EFL_AF \| X86_EFL_PF)
1562	static const struct
1563	{
1564	RTCCUINTREG uInDX;
1565	RTCCUINTREG uInAX;
1566	RTCCUINTREG uInBX;
1567	RTCCUINTREG uOutAX;
1568	RTCCUINTREG uOutDX;
1569	uint8_t bXcpt;
1570	} s_aTests[] =
1571	{
1572	{ 0, 0, 0,
1573	0, 0, X86_XCPT_DE },
1574	{ RTCCINTREG_MAX, RTCCINTREG_MAX, 0,
1575	0, 0, X86_XCPT_DE },
1576	/* two positive values. */
1577	{ 0, 1, 1,
1578	1, 0, X86_XCPT_UD },
1579	{ 0, 5, 2,
1580	2, 1, X86_XCPT_UD },
1581	{ RTCCINTREG_MAX / 2, RTCCUINTREG_MAX / 2, RTCCINTREG_MAX,
1582	RTCCINTREG_MAX, RTCCINTREG_MAX - 1, X86_XCPT_UD },
1583	{ RTCCINTREG_MAX / 2, RTCCUINTREG_MAX / 2 + 1, RTCCINTREG_MAX,
1584	RTCCINTREG_MAX, RTCCINTREG_MAX - 1, X86_XCPT_DE },
1585	/* negative dividend, positive divisor. */
1586	{ -1, -7, 2,
1587	-3, -1, X86_XCPT_UD },
1588	{ RTCCINTREG_MIN / 2 + 1, 0, RTCCINTREG_MAX,
1589	RTCCINTREG_MIN + 2, RTCCINTREG_MIN + 2, X86_XCPT_UD },
1590	{ RTCCINTREG_MIN / 2, 0, RTCCINTREG_MAX,
1591	0, 0, X86_XCPT_DE },
1592	/* positive dividend, negative divisor. */
1593	{ 0, 7, -2,
1594	-3, 1, X86_XCPT_UD },
1595	{ RTCCINTREG_MAX / 2 + 1, RTCCINTREG_MAX, RTCCINTREG_MIN,
1596	RTCCINTREG_MIN, RTCCINTREG_MAX, X86_XCPT_UD },
1597	{ RTCCINTREG_MAX / 2 + 1, (RTCCUINTREG)RTCCINTREG_MAX+1, RTCCINTREG_MIN,
1598	0, 0, X86_XCPT_DE },
1599	/* negative dividend, negative divisor. */
1600	{ -1, -7, -2,
1601	3, -1, X86_XCPT_UD },
1602	{ RTCCINTREG_MIN / 2, 1, RTCCINTREG_MIN,
1603	RTCCINTREG_MAX, RTCCINTREG_MIN + 1, X86_XCPT_UD },
1604	{ RTCCINTREG_MIN / 2, 2, RTCCINTREG_MIN,
1605	RTCCINTREG_MAX, RTCCINTREG_MIN + 2, X86_XCPT_UD },
1606	{ RTCCINTREG_MIN / 2, 0, RTCCINTREG_MIN,
1607	0, 0, X86_XCPT_DE },
1608	};
1609
1610	BS3REGCTX Ctx;
1611	BS3TRAPFRAME TrapFrame;
1612	unsigned i, j;
1613
1614	/* Ensure the structures are allocated before we sample the stack pointer. */
1615	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
1616	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
1617
1618	/*
1619	* Create test context.
1620	*/
1621	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
1622	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, BS3_CMN_NM(bs3CpuInstr2_idiv_xBX_ud2));
1623
1624	/*
1625	* Do the tests twice, first with all flags set, then once again with
1626	* flags cleared. The flags are not touched by my intel skylake CPU.
1627	*/
1628	Ctx.rflags.u16 \|= IDIV_CHECK_EFLAGS;
1629	for (j = 0; j < 2; j++)
1630	{
1631	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
1632	{
1633	Ctx.rax.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInAX;
1634	Ctx.rdx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInDX;
1635	Ctx.rbx.RT_CONCAT(u,ARCH_BITS) = s_aTests[i].uInBX;
1636	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
1637
1638	if ( TrapFrame.bXcpt != s_aTests[i].bXcpt
1639	\|\| ( s_aTests[i].bXcpt == X86_XCPT_UD
1640	? TrapFrame.Ctx.rax.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutAX
1641	\|\| TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutDX
1642	\|\| (TrapFrame.Ctx.rflags.u16 & IDIV_CHECK_EFLAGS) != (Ctx.rflags.u16 & IDIV_CHECK_EFLAGS)
1643	: TrapFrame.Ctx.rax.u != Ctx.rax.u
1644	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
1645	\|\| (TrapFrame.Ctx.rflags.u16 & IDIV_CHECK_EFLAGS) != (Ctx.rflags.u16 & IDIV_CHECK_EFLAGS) ) )
1646	{
1647	Bs3TestFailedF("test #%i failed: input %#" RTCCUINTREG_XFMT ":%" RTCCUINTREG_XFMT " / %#" RTCCUINTREG_XFMT,
1648	i, s_aTests[i].uInDX, s_aTests[i].uInAX, s_aTests[i].uInBX);
1649	if (TrapFrame.bXcpt != s_aTests[i].bXcpt)
1650	Bs3TestFailedF("Expected bXcpt = %#x, got %#x", s_aTests[i].bXcpt, TrapFrame.bXcpt);
1651	if (s_aTests[i].bXcpt == X86_XCPT_UD)
1652	{
1653	if (TrapFrame.Ctx.rax.RT_CONCAT(u, ARCH_BITS) != s_aTests[i].uOutAX)
1654	Bs3TestFailedF("Expected xAX = %#" RTCCUINTREG_XFMT ", got %#" RTCCUINTREG_XFMT,
1655	s_aTests[i].uOutAX, TrapFrame.Ctx.rax.RT_CONCAT(u,ARCH_BITS));
1656	if (TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS) != s_aTests[i].uOutDX)
1657	Bs3TestFailedF("Expected xDX = %#" RTCCUINTREG_XFMT ", got %#" RTCCUINTREG_XFMT,
1658	s_aTests[i].uOutDX, TrapFrame.Ctx.rdx.RT_CONCAT(u,ARCH_BITS));
1659	if ((TrapFrame.Ctx.rflags.u16 & IDIV_CHECK_EFLAGS) != (Ctx.rflags.u16 & IDIV_CHECK_EFLAGS))
1660	Bs3TestFailedF("Expected EFLAGS = %#06RX16, got %#06RX16",
1661	Ctx.rflags.u16 & IDIV_CHECK_EFLAGS, TrapFrame.Ctx.rflags.u16 & IDIV_CHECK_EFLAGS);
1662	}
1663	}
1664	}
1665	Ctx.rflags.u16 &= ~IDIV_CHECK_EFLAGS;
1666	}
1667
1668	return 0;
1669	}
1670
1671
1672	/*
1673	* BSF/BSR (386+) & TZCNT/LZCNT (BMI1,ABM)
1674	*/
1675
1676	typedef struct BS3CPUINSTR2_SUBTEST_BITSCAN_T
1677	{
1678	RTCCUINTXREG uSrc;
1679	RTCCUINTXREG uOut;
1680	bool fOutNotSet;
1681	uint16_t fEflOut;
1682	} BS3CPUINSTR2_SUBTEST_BITSCAN_T;
1683
1684	typedef struct BS3CPUINSTR2_TEST_BITSCAN_T
1685	{
1686	FPFNBS3FAR pfnWorker;
1687	bool fMemSrc;
1688	uint8_t cbInstr;
1689	uint8_t cOpBits;
1690	uint16_t fEflCheck;
1691	uint8_t cSubTests;
1692	BS3CPUINSTR2_SUBTEST_BITSCAN_T const *paSubTests;
1693	} BS3CPUINSTR2_TEST_BITSCAN_T;
1694
1695	static uint8_t bs3CpuInstr2_BitScan(uint8_t bMode, BS3CPUINSTR2_TEST_BITSCAN_T const *paTests, unsigned cTests)
1696	{
1697	BS3REGCTX Ctx;
1698	BS3TRAPFRAME TrapFrame;
1699	unsigned i, j, k;
1700
1701	/* Ensure the structures are allocated before we sample the stack pointer. */
1702	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
1703	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
1704
1705	/*
1706	* Create test context.
1707	*/
1708	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
1709
1710	/*
1711	* Do the tests twice, first with all flags set, then once again with
1712	* flags cleared. The flags are not supposed to be touched at all.
1713	*/
1714	Ctx.rflags.u16 \|= X86_EFL_STATUS_BITS;
1715	for (j = 0; j < 2; j++)
1716	{
1717	for (i = 0; i < cTests; i++)
1718	{
1719	for (k = 0; k < paTests[i].cSubTests; k++)
1720	{
1721	uint64_t uExpectRax, uExpectRip;
1722	RTCCUINTXREG uMemSrc, uMemSrcExpect;
1723
1724	Ctx.rax.uCcXReg = RTCCUINTXREG_MAX * 1019;
1725	if (!paTests[i].fMemSrc)
1726	{
1727	Ctx.rbx.uCcXReg = paTests[i].paSubTests[k].uSrc;
1728	uMemSrcExpect = uMemSrc = ~paTests[i].paSubTests[k].uSrc;
1729	}
1730	else
1731	{
1732	uMemSrcExpect = uMemSrc = paTests[i].paSubTests[k].uSrc;
1733	Bs3RegCtxSetGrpSegFromCurPtr(&Ctx, &Ctx.rbx, &Ctx.fs, &uMemSrc);
1734	}
1735	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, paTests[i].pfnWorker);
1736	if (paTests[i].paSubTests[k].fOutNotSet)
1737	uExpectRax = Ctx.rax.u;
1738	else if (paTests[i].cOpBits != 16)
1739	uExpectRax = paTests[i].paSubTests[k].uOut;
1740	else
1741	uExpectRax = paTests[i].paSubTests[k].uOut \| (Ctx.rax.u & UINT64_C(0xffffffffffff0000));
1742	uExpectRip = Ctx.rip.u + paTests[i].cbInstr;
1743	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
1744
1745	if ( TrapFrame.bXcpt != X86_XCPT_UD
1746	\|\| TrapFrame.Ctx.rip.u != uExpectRip
1747	\|\| TrapFrame.Ctx.rbx.u != Ctx.rbx.u
1748	\|\| TrapFrame.Ctx.rax.u != uExpectRax
1749	\|\| (TrapFrame.Ctx.rflags.u16 & paTests[i].fEflCheck)
1750	!= (paTests[i].paSubTests[k].fEflOut & paTests[i].fEflCheck)
1751	/* check that nothing else really changed: */
1752	\|\| TrapFrame.Ctx.rcx.u != Ctx.rcx.u
1753	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
1754	\|\| TrapFrame.Ctx.rsp.u != Ctx.rsp.u
1755	\|\| TrapFrame.Ctx.rbp.u != Ctx.rbp.u
1756	\|\| TrapFrame.Ctx.rsi.u != Ctx.rsi.u
1757	\|\| TrapFrame.Ctx.rdi.u != Ctx.rdi.u
1758	\|\| uMemSrc != uMemSrcExpect
1759	)
1760	{
1761	Bs3TestFailedF("test #%i/%i failed: input %#" RTCCUINTXREG_XFMT,
1762	i, k, paTests[i].paSubTests[k].uSrc);
1763	if (TrapFrame.bXcpt != X86_XCPT_UD)
1764	Bs3TestFailedF("Expected bXcpt = %#x, got %#x", X86_XCPT_UD, TrapFrame.bXcpt);
1765	if (TrapFrame.Ctx.rip.u != uExpectRip)
1766	Bs3TestFailedF("Expected RIP = %#06RX64, got %#06RX64", uExpectRip, TrapFrame.Ctx.rip.u);
1767	if (TrapFrame.Ctx.rax.u != uExpectRax)
1768	Bs3TestFailedF("Expected RAX = %#06RX64, got %#06RX64", uExpectRax, TrapFrame.Ctx.rax.u);
1769	if (TrapFrame.Ctx.rcx.u != Ctx.rcx.u)
1770	Bs3TestFailedF("Expected RCX = %#06RX64, got %#06RX64", Ctx.rcx.u, TrapFrame.Ctx.rcx.u);
1771	if (TrapFrame.Ctx.rbx.u != Ctx.rbx.u)
1772	Bs3TestFailedF("Expected RBX = %#06RX64, got %#06RX64 (dst)", Ctx.rbx.u, TrapFrame.Ctx.rbx.u);
1773	if ( (TrapFrame.Ctx.rflags.u16 & paTests[i].fEflCheck)
1774	!= (paTests[i].paSubTests[k].fEflOut & paTests[i].fEflCheck))
1775	Bs3TestFailedF("Expected EFLAGS = %#06RX32, got %#06RX32 (output)",
1776	paTests[i].paSubTests[k].fEflOut & paTests[i].fEflCheck,
1777	TrapFrame.Ctx.rflags.u16 & paTests[i].fEflCheck);
1778
1779	if (TrapFrame.Ctx.rdx.u != Ctx.rdx.u)
1780	Bs3TestFailedF("Expected RDX = %#06RX64, got %#06RX64 (src)", Ctx.rdx.u, TrapFrame.Ctx.rdx.u);
1781	if (TrapFrame.Ctx.rsp.u != Ctx.rsp.u)
1782	Bs3TestFailedF("Expected RSP = %#06RX64, got %#06RX64", Ctx.rsp.u, TrapFrame.Ctx.rsp.u);
1783	if (TrapFrame.Ctx.rbp.u != Ctx.rbp.u)
1784	Bs3TestFailedF("Expected RBP = %#06RX64, got %#06RX64", Ctx.rbp.u, TrapFrame.Ctx.rbp.u);
1785	if (TrapFrame.Ctx.rsi.u != Ctx.rsi.u)
1786	Bs3TestFailedF("Expected RSI = %#06RX64, got %#06RX64", Ctx.rsi.u, TrapFrame.Ctx.rsi.u);
1787	if (TrapFrame.Ctx.rdi.u != Ctx.rdi.u)
1788	Bs3TestFailedF("Expected RDI = %#06RX64, got %#06RX64", Ctx.rdi.u, TrapFrame.Ctx.rdi.u);
1789	if (uMemSrc != uMemSrcExpect)
1790	Bs3TestFailedF("Expected uMemSrc = %#06RX64, got %#06RX64", (uint64_t)uMemSrcExpect, (uint64_t)uMemSrc);
1791	}
1792	}
1793	}
1794	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
1795	}
1796
1797	return 0;
1798	}
1799
1800
1801	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_bsf_tzcnt)(uint8_t bMode)
1802	{
1803	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsBsf16[] =
1804	{
1805	{ 0, /* -> */ 0, true, X86_EFL_ZF },
1806	{ ~(RTCCUINTXREG)UINT16_MAX, /* -> */ 0, true, X86_EFL_ZF },
1807	{ ~(RTCCUINTXREG)0, /* -> */ 0, false, 0 },
1808	{ ~(RTCCUINTXREG)1, /* -> */ 1, false, 0 },
1809	{ UINT16_C(0x8000), /* -> */ 15, false, 0 },
1810	{ UINT16_C(0x4560), /* -> */ 5, false, 0 },
1811	};
1812	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsTzCnt16[] =
1813	{
1814	{ 0, /* -> */ 16, false, X86_EFL_CF },
1815	{ ~(RTCCUINTXREG)UINT16_MAX, /* -> */ 16, false, X86_EFL_CF },
1816	{ ~(RTCCUINTXREG)0, /* -> */ 0, false, X86_EFL_ZF },
1817	{ ~(RTCCUINTXREG)1, /* -> */ 1, false, 0 },
1818	{ UINT16_C(0x8000), /* -> */ 15, false, 0 },
1819	{ UINT16_C(0x4560), /* -> */ 5, false, 0 },
1820	};
1821	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsBsf32[] =
1822	{
1823	{ 0, /* -> */ 0, true, X86_EFL_ZF },
1824	#if ARCH_BITS == 64
1825	{ ~(RTCCUINTXREG)UINT32_MAX, /* -> */ 0, true, X86_EFL_ZF },
1826	#endif
1827	{ ~(RTCCUINTXREG)0, /* -> */ 0, false, 0 },
1828	{ ~(RTCCUINTXREG)1, /* -> */ 1, false, 0 },
1829	{ UINT16_C(0x8000), /* -> */ 15, false, 0 },
1830	{ UINT16_C(0x4560), /* -> */ 5, false, 0 },
1831	{ UINT32_C(0x80000000), /* -> */ 31, false, 0 },
1832	{ UINT32_C(0x45600000), /* -> */ 21, false, 0 },
1833	};
1834	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsTzCnt32[] =
1835	{
1836	{ 0, /* -> */ 32, false, X86_EFL_CF },
1837	#if ARCH_BITS == 64
1838	{ ~(RTCCUINTXREG)UINT32_MAX, /* -> */ 32, false, X86_EFL_CF },
1839	#endif
1840	{ ~(RTCCUINTXREG)0, /* -> */ 0, false, X86_EFL_ZF },
1841	{ ~(RTCCUINTXREG)1, /* -> */ 1, false, 0 },
1842	{ UINT16_C(0x8000), /* -> */ 15, false, 0 },
1843	{ UINT16_C(0x4560), /* -> */ 5, false, 0 },
1844	{ UINT32_C(0x80000000), /* -> */ 31, false, 0 },
1845	{ UINT32_C(0x45600000), /* -> */ 21, false, 0 },
1846	};
1847	#if ARCH_BITS == 64
1848	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsBsf64[] =
1849	{
1850	{ 0, /* -> */ 0, true, X86_EFL_ZF },
1851	{ ~(RTCCUINTXREG)0, /* -> */ 0, false, 0 },
1852	{ ~(RTCCUINTXREG)1, /* -> */ 1, false, 0 },
1853	{ UINT16_C(0x8000), /* -> */ 15, false, 0 },
1854	{ UINT16_C(0x4560), /* -> */ 5, false, 0 },
1855	{ UINT32_C(0x80000000), /* -> */ 31, false, 0 },
1856	{ UINT32_C(0x45600000), /* -> */ 21, false, 0 },
1857	{ UINT64_C(0x8000000000000000), /* -> */ 63, false, 0 },
1858	{ UINT64_C(0x4560000000000000), /* -> */ 53, false, 0 },
1859	};
1860	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsTzCnt64[] =
1861	{
1862	{ 0, /* -> */ 64, false, X86_EFL_CF },
1863	{ ~(RTCCUINTXREG)0, /* -> */ 0, false, X86_EFL_ZF },
1864	{ ~(RTCCUINTXREG)1, /* -> */ 1, false, 0 },
1865	{ UINT16_C(0x8000), /* -> */ 15, false, 0 },
1866	{ UINT16_C(0x4560), /* -> */ 5, false, 0 },
1867	{ UINT32_C(0x80000000), /* -> */ 31, false, 0 },
1868	{ UINT32_C(0x45600000), /* -> */ 21, false, 0 },
1869	{ UINT64_C(0x8000000000000000), /* -> */ 63, false, 0 },
1870	{ UINT64_C(0x4560000000000000), /* -> */ 53, false, 0 },
1871	};
1872	#endif
1873	static BS3CPUINSTR2_TEST_BITSCAN_T s_aTests[] =
1874	{
1875	{ BS3_CMN_NM(bs3CpuInstr2_bsf_AX_BX_ud2), false, 3 + (ARCH_BITS != 16), 16, X86_EFL_ZF,
1876	RT_ELEMENTS(s_aSubTestsBsf16), s_aSubTestsBsf16 },
1877	{ BS3_CMN_NM(bs3CpuInstr2_bsf_AX_FSxBX_ud2), true, 4 + (ARCH_BITS != 16), 16, X86_EFL_ZF,
1878	RT_ELEMENTS(s_aSubTestsBsf16), s_aSubTestsBsf16 },
1879	{ BS3_CMN_NM(bs3CpuInstr2_bsf_EAX_EBX_ud2), false, 3 + (ARCH_BITS == 16), 32, X86_EFL_ZF,
1880	RT_ELEMENTS(s_aSubTestsBsf32), s_aSubTestsBsf32 },
1881	{ BS3_CMN_NM(bs3CpuInstr2_bsf_EAX_FSxBX_ud2), true, 4 + (ARCH_BITS == 16), 32, X86_EFL_ZF,
1882	RT_ELEMENTS(s_aSubTestsBsf32), s_aSubTestsBsf32 },
1883	#if ARCH_BITS == 64
1884	{ BS3_CMN_NM(bs3CpuInstr2_bsf_RAX_RBX_ud2), false, 4, 64, X86_EFL_ZF,
1885	RT_ELEMENTS(s_aSubTestsBsf64), s_aSubTestsBsf64 },
1886	{ BS3_CMN_NM(bs3CpuInstr2_bsf_RAX_FSxBX_ud2), true, 5, 64, X86_EFL_ZF,
1887	RT_ELEMENTS(s_aSubTestsBsf64), s_aSubTestsBsf64 },
1888	#endif
1889	/* f2 prefixed variant: */
1890	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsf_AX_BX_ud2), false, 4 + (ARCH_BITS != 16), 16, X86_EFL_ZF,
1891	RT_ELEMENTS(s_aSubTestsBsf16), s_aSubTestsBsf16 },
1892	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsf_AX_FSxBX_ud2), true, 5 + (ARCH_BITS != 16), 16, X86_EFL_ZF,
1893	RT_ELEMENTS(s_aSubTestsBsf16), s_aSubTestsBsf16 },
1894	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsf_EAX_EBX_ud2), false, 4 + (ARCH_BITS == 16), 32, X86_EFL_ZF,
1895	RT_ELEMENTS(s_aSubTestsBsf32), s_aSubTestsBsf32 },
1896	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsf_EAX_FSxBX_ud2), true, 5 + (ARCH_BITS == 16), 32, X86_EFL_ZF,
1897	RT_ELEMENTS(s_aSubTestsBsf32), s_aSubTestsBsf32 },
1898	#if ARCH_BITS == 64
1899	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsf_RAX_RBX_ud2), false, 5, 64, X86_EFL_ZF,
1900	RT_ELEMENTS(s_aSubTestsBsf64), s_aSubTestsBsf64 },
1901	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsf_RAX_FSxBX_ud2), true, 6, 64, X86_EFL_ZF,
1902	RT_ELEMENTS(s_aSubTestsBsf64), s_aSubTestsBsf64 },
1903	#endif
1904
1905	/* tzcnt: */
1906	{ BS3_CMN_NM(bs3CpuInstr2_tzcnt_AX_BX_ud2), false, 4 + (ARCH_BITS != 16), 16, X86_EFL_ZF \| X86_EFL_CF,
1907	RT_ELEMENTS(s_aSubTestsTzCnt16), s_aSubTestsTzCnt16 },
1908	{ BS3_CMN_NM(bs3CpuInstr2_tzcnt_AX_FSxBX_ud2), true, 5 + (ARCH_BITS != 16), 16, X86_EFL_ZF \| X86_EFL_CF,
1909	RT_ELEMENTS(s_aSubTestsTzCnt16), s_aSubTestsTzCnt16 },
1910	{ BS3_CMN_NM(bs3CpuInstr2_tzcnt_EAX_EBX_ud2), false, 4 + (ARCH_BITS == 16), 32, X86_EFL_ZF \| X86_EFL_CF,
1911	RT_ELEMENTS(s_aSubTestsTzCnt32), s_aSubTestsTzCnt32 },
1912	{ BS3_CMN_NM(bs3CpuInstr2_tzcnt_EAX_FSxBX_ud2), true, 5 + (ARCH_BITS == 16), 32, X86_EFL_ZF \| X86_EFL_CF,
1913	RT_ELEMENTS(s_aSubTestsTzCnt32), s_aSubTestsTzCnt32 },
1914	#if ARCH_BITS == 64
1915	{ BS3_CMN_NM(bs3CpuInstr2_tzcnt_RAX_RBX_ud2), false, 5, 64, X86_EFL_ZF \| X86_EFL_CF,
1916	RT_ELEMENTS(s_aSubTestsTzCnt64), s_aSubTestsTzCnt64 },
1917	{ BS3_CMN_NM(bs3CpuInstr2_tzcnt_RAX_FSxBX_ud2), true, 6, 64, X86_EFL_ZF \| X86_EFL_CF,
1918	RT_ELEMENTS(s_aSubTestsTzCnt64), s_aSubTestsTzCnt64 },
1919	#endif
1920	/* f2 prefixed tzcnt variant (last prefix (f3) should prevail): */
1921	{ BS3_CMN_NM(bs3CpuInstr2_f2_tzcnt_AX_BX_ud2), false, 5 + (ARCH_BITS != 16), 16, X86_EFL_ZF \| X86_EFL_CF,
1922	RT_ELEMENTS(s_aSubTestsTzCnt16), s_aSubTestsTzCnt16 },
1923	{ BS3_CMN_NM(bs3CpuInstr2_f2_tzcnt_AX_FSxBX_ud2), true, 6 + (ARCH_BITS != 16), 16, X86_EFL_ZF \| X86_EFL_CF,
1924	RT_ELEMENTS(s_aSubTestsTzCnt16), s_aSubTestsTzCnt16 },
1925	{ BS3_CMN_NM(bs3CpuInstr2_f2_tzcnt_EAX_EBX_ud2), false, 5 + (ARCH_BITS == 16), 32, X86_EFL_ZF \| X86_EFL_CF,
1926	RT_ELEMENTS(s_aSubTestsTzCnt32), s_aSubTestsTzCnt32 },
1927	{ BS3_CMN_NM(bs3CpuInstr2_f2_tzcnt_EAX_FSxBX_ud2),true, 6 + (ARCH_BITS == 16), 32, X86_EFL_ZF \| X86_EFL_CF,
1928	RT_ELEMENTS(s_aSubTestsTzCnt32), s_aSubTestsTzCnt32 },
1929	#if ARCH_BITS == 64
1930	{ BS3_CMN_NM(bs3CpuInstr2_f2_tzcnt_RAX_RBX_ud2), false, 6, 64, X86_EFL_ZF \| X86_EFL_CF,
1931	RT_ELEMENTS(s_aSubTestsTzCnt64), s_aSubTestsTzCnt64 },
1932	{ BS3_CMN_NM(bs3CpuInstr2_f2_tzcnt_RAX_FSxBX_ud2),true, 7, 64, X86_EFL_ZF \| X86_EFL_CF,
1933	RT_ELEMENTS(s_aSubTestsTzCnt64), s_aSubTestsTzCnt64 },
1934	#endif
1935	};
1936
1937	uint32_t uStdExtFeatEbx = 0;
1938	if (g_uBs3CpuDetected & BS3CPU_F_CPUID)
1939	ASMCpuIdExSlow(7, 0, 0, 0, NULL, &uStdExtFeatEbx, NULL, NULL);
1940	if (!(uStdExtFeatEbx & X86_CPUID_STEXT_FEATURE_EBX_BMI1))
1941	{
1942	unsigned i = RT_ELEMENTS(s_aTests);
1943	while (i-- > 0)
1944	if (s_aTests[i].fEflCheck & X86_EFL_CF)
1945	{
1946	s_aTests[i].fEflCheck = X86_EFL_ZF;
1947	switch (s_aTests[i].cOpBits)
1948	{
1949	case 16:
1950	s_aTests[i].cSubTests = RT_ELEMENTS(s_aSubTestsBsf16);
1951	s_aTests[i].paSubTests = s_aSubTestsBsf16;
1952	break;
1953	case 32:
1954	s_aTests[i].cSubTests = RT_ELEMENTS(s_aSubTestsBsf32);
1955	s_aTests[i].paSubTests = s_aSubTestsBsf32;
1956	break;
1957	#if ARCH_BITS == 64
1958	case 64:
1959	s_aTests[i].cSubTests = RT_ELEMENTS(s_aSubTestsBsf64);
1960	s_aTests[i].paSubTests = s_aSubTestsBsf64;
1961	break;
1962	#endif
1963	}
1964	}
1965	Bs3TestPrintf("tzcnt not supported\n");
1966	}
1967
1968	return bs3CpuInstr2_BitScan(bMode, s_aTests, RT_ELEMENTS(s_aTests));
1969	}
1970
1971
1972	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_bsr_lzcnt)(uint8_t bMode)
1973	{
1974	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsBsr16[] =
1975	{
1976	{ 0, /* -> */ 0, true, X86_EFL_ZF },
1977	{ ~(RTCCUINTXREG)UINT16_MAX, /* -> */ 0, true, X86_EFL_ZF },
1978	{ ~(RTCCUINTXREG)0, /* -> */ 15, false, 0 },
1979	{ ~(RTCCUINTXREG)1, /* -> */ 15, false, 0 },
1980	{ UINT16_C(0x0001), /* -> */ 0, false, 0 },
1981	{ UINT16_C(0x0002), /* -> */ 1, false, 0 },
1982	{ UINT16_C(0x4560), /* -> */ 14, false, 0 },
1983	};
1984	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsLzCnt16[] =
1985	{
1986	{ 0, /* -> */ 16, false, X86_EFL_CF },
1987	{ ~(RTCCUINTXREG)UINT16_MAX, /* -> */ 16, false, X86_EFL_CF },
1988	{ ~(RTCCUINTXREG)0, /* -> */ 0, false, X86_EFL_ZF },
1989	{ ~(RTCCUINTXREG)1, /* -> */ 0, false, X86_EFL_ZF },
1990	{ UINT16_C(0x8000), /* -> */ 0, false, X86_EFL_ZF },
1991	{ UINT16_C(0x4560), /* -> */ 1, false, 0 },
1992	{ UINT16_C(0x003f), /* -> */ 10, false, 0 },
1993	{ UINT16_C(0x0001), /* -> */ 15, false, 0 },
1994	};
1995	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsBsr32[] =
1996	{
1997	{ 0, /* -> */ 0, true, X86_EFL_ZF },
1998	#if ARCH_BITS == 64
1999	{ ~(RTCCUINTXREG)UINT32_MAX, /* -> */ 0, true, X86_EFL_ZF },
2000	#endif
2001	{ ~(RTCCUINTXREG)0, /* -> */ 31, false, 0 },
2002	{ ~(RTCCUINTXREG)1, /* -> */ 31, false, 0 },
2003	{ 1, /* -> */ 0, false, 0 },
2004	{ 2, /* -> */ 1, false, 0 },
2005	{ UINT16_C(0x8000), /* -> */ 15, false, 0 },
2006	{ UINT16_C(0x4560), /* -> */ 14, false, 0 },
2007	{ UINT32_C(0x80000000), /* -> */ 31, false, 0 },
2008	{ UINT32_C(0x45600000), /* -> */ 30, false, 0 },
2009	};
2010	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsLzCnt32[] =
2011	{
2012	{ 0, /* -> */ 32, false, X86_EFL_CF },
2013	#if ARCH_BITS == 64
2014	{ ~(RTCCUINTXREG)UINT32_MAX, /* -> */ 32, false, X86_EFL_CF },
2015	#endif
2016	{ ~(RTCCUINTXREG)0, /* -> */ 0, false, X86_EFL_ZF },
2017	{ ~(RTCCUINTXREG)1, /* -> */ 0, false, X86_EFL_ZF },
2018	{ 1, /* -> */ 31, false, 0 },
2019	{ 2, /* -> */ 30, false, 0},
2020	{ UINT16_C(0x8000), /* -> */ 16, false, 0 },
2021	{ UINT16_C(0x4560), /* -> */ 17, false, 0 },
2022	{ UINT32_C(0x80000000), /* -> */ 0, false, X86_EFL_ZF },
2023	{ UINT32_C(0x45600000), /* -> */ 1, false, 0 },
2024	{ UINT32_C(0x0000ffff), /* -> */ 16, false, 0 },
2025	};
2026	#if ARCH_BITS == 64
2027	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsBsr64[] =
2028	{
2029	{ 0, /* -> */ 0, true, X86_EFL_ZF },
2030	{ ~(RTCCUINTXREG)0, /* -> */ 63, false, 0 },
2031	{ ~(RTCCUINTXREG)1, /* -> */ 63, false, 0 },
2032	{ 1, /* -> */ 0, false, 0 },
2033	{ 2, /* -> */ 1, false, 0 },
2034	{ UINT16_C(0x8000), /* -> */ 15, false, 0 },
2035	{ UINT16_C(0x4560), /* -> */ 14, false, 0 },
2036	{ UINT32_C(0x80000000), /* -> */ 31, false, 0 },
2037	{ UINT32_C(0x45600000), /* -> */ 30, false, 0 },
2038	{ UINT64_C(0x8000000000000000), /* -> */ 63, false, 0 },
2039	{ UINT64_C(0x0045600000000000), /* -> */ 54, false, 0 },
2040	};
2041	static BS3CPUINSTR2_SUBTEST_BITSCAN_T const s_aSubTestsLzCnt64[] =
2042	{
2043	{ 0, /* -> */ 64, false, X86_EFL_CF },
2044	{ ~(RTCCUINTXREG)0, /* -> */ 0, false, X86_EFL_ZF },
2045	{ ~(RTCCUINTXREG)1, /* -> */ 0, false, X86_EFL_ZF },
2046	{ 1, /* -> */ 63, false, 0 },
2047	{ 2, /* -> */ 62, false, 0 },
2048	{ UINT16_C(0x8000), /* -> */ 48, false, 0 },
2049	{ UINT16_C(0x4560), /* -> */ 49, false, 0 },
2050	{ UINT32_C(0x80000000), /* -> */ 32, false, 0 },
2051	{ UINT32_C(0x45600000), /* -> */ 33, false, 0 },
2052	{ UINT64_C(0x8000000000000000), /* -> */ 0, false, X86_EFL_ZF },
2053	{ UINT64_C(0x4560000000000000), /* -> */ 1, false, 0 },
2054	{ UINT64_C(0x0045600000000000), /* -> */ 9, false, 0 },
2055	};
2056	#endif
2057	static BS3CPUINSTR2_TEST_BITSCAN_T s_aTests[] =
2058	{
2059	{ BS3_CMN_NM(bs3CpuInstr2_bsr_AX_BX_ud2), false, 3 + (ARCH_BITS != 16), 16, X86_EFL_ZF,
2060	RT_ELEMENTS(s_aSubTestsBsr16), s_aSubTestsBsr16 },
2061	{ BS3_CMN_NM(bs3CpuInstr2_bsr_AX_FSxBX_ud2), true, 4 + (ARCH_BITS != 16), 16, X86_EFL_ZF,
2062	RT_ELEMENTS(s_aSubTestsBsr16), s_aSubTestsBsr16 },
2063	{ BS3_CMN_NM(bs3CpuInstr2_bsr_EAX_EBX_ud2), false, 3 + (ARCH_BITS == 16), 32, X86_EFL_ZF,
2064	RT_ELEMENTS(s_aSubTestsBsr32), s_aSubTestsBsr32 },
2065	{ BS3_CMN_NM(bs3CpuInstr2_bsr_EAX_FSxBX_ud2), true, 4 + (ARCH_BITS == 16), 32, X86_EFL_ZF,
2066	RT_ELEMENTS(s_aSubTestsBsr32), s_aSubTestsBsr32 },
2067	#if ARCH_BITS == 64
2068	{ BS3_CMN_NM(bs3CpuInstr2_bsr_RAX_RBX_ud2), false, 4, 64, X86_EFL_ZF,
2069	RT_ELEMENTS(s_aSubTestsBsr64), s_aSubTestsBsr64 },
2070	{ BS3_CMN_NM(bs3CpuInstr2_bsr_RAX_FSxBX_ud2), true, 5, 64, X86_EFL_ZF,
2071	RT_ELEMENTS(s_aSubTestsBsr64), s_aSubTestsBsr64 },
2072	#endif
2073	/* f2 prefixed variant: */
2074	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsr_AX_BX_ud2), false, 4 + (ARCH_BITS != 16), 16, X86_EFL_ZF,
2075	RT_ELEMENTS(s_aSubTestsBsr16), s_aSubTestsBsr16 },
2076	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsr_AX_FSxBX_ud2), true, 5 + (ARCH_BITS != 16), 16, X86_EFL_ZF,
2077	RT_ELEMENTS(s_aSubTestsBsr16), s_aSubTestsBsr16 },
2078	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsr_EAX_EBX_ud2), false, 4 + (ARCH_BITS == 16), 32, X86_EFL_ZF,
2079	RT_ELEMENTS(s_aSubTestsBsr32), s_aSubTestsBsr32 },
2080	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsr_EAX_FSxBX_ud2), true, 5 + (ARCH_BITS == 16), 32, X86_EFL_ZF,
2081	RT_ELEMENTS(s_aSubTestsBsr32), s_aSubTestsBsr32 },
2082	#if ARCH_BITS == 64
2083	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsr_RAX_RBX_ud2), false, 5, 64, X86_EFL_ZF,
2084	RT_ELEMENTS(s_aSubTestsBsr64), s_aSubTestsBsr64 },
2085	{ BS3_CMN_NM(bs3CpuInstr2_f2_bsr_RAX_FSxBX_ud2), true, 6, 64, X86_EFL_ZF,
2086	RT_ELEMENTS(s_aSubTestsBsr64), s_aSubTestsBsr64 },
2087	#endif
2088
2089	/* lzcnt: */
2090	{ BS3_CMN_NM(bs3CpuInstr2_lzcnt_AX_BX_ud2), false, 4 + (ARCH_BITS != 16), 16, X86_EFL_ZF \| X86_EFL_CF,
2091	RT_ELEMENTS(s_aSubTestsLzCnt16), s_aSubTestsLzCnt16 },
2092	{ BS3_CMN_NM(bs3CpuInstr2_lzcnt_AX_FSxBX_ud2), true, 5 + (ARCH_BITS != 16), 16, X86_EFL_ZF \| X86_EFL_CF,
2093	RT_ELEMENTS(s_aSubTestsLzCnt16), s_aSubTestsLzCnt16 },
2094	{ BS3_CMN_NM(bs3CpuInstr2_lzcnt_EAX_EBX_ud2), false, 4 + (ARCH_BITS == 16), 32, X86_EFL_ZF \| X86_EFL_CF,
2095	RT_ELEMENTS(s_aSubTestsLzCnt32), s_aSubTestsLzCnt32 },
2096	{ BS3_CMN_NM(bs3CpuInstr2_lzcnt_EAX_FSxBX_ud2), true, 5 + (ARCH_BITS == 16), 32, X86_EFL_ZF \| X86_EFL_CF,
2097	RT_ELEMENTS(s_aSubTestsLzCnt32), s_aSubTestsLzCnt32 },
2098	#if ARCH_BITS == 64
2099	{ BS3_CMN_NM(bs3CpuInstr2_lzcnt_RAX_RBX_ud2), false, 5, 64, X86_EFL_ZF \| X86_EFL_CF,
2100	RT_ELEMENTS(s_aSubTestsLzCnt64), s_aSubTestsLzCnt64 },
2101	{ BS3_CMN_NM(bs3CpuInstr2_lzcnt_RAX_FSxBX_ud2), true, 6, 64, X86_EFL_ZF \| X86_EFL_CF,
2102	RT_ELEMENTS(s_aSubTestsLzCnt64), s_aSubTestsLzCnt64 },
2103	#endif
2104	/* f2 prefixed lzcnt variant (last prefix (f3) should prevail): */
2105	{ BS3_CMN_NM(bs3CpuInstr2_f2_lzcnt_AX_BX_ud2), false, 5 + (ARCH_BITS != 16), 16, X86_EFL_ZF \| X86_EFL_CF,
2106	RT_ELEMENTS(s_aSubTestsLzCnt16), s_aSubTestsLzCnt16 },
2107	{ BS3_CMN_NM(bs3CpuInstr2_f2_lzcnt_AX_FSxBX_ud2), true, 6 + (ARCH_BITS != 16), 16, X86_EFL_ZF \| X86_EFL_CF,
2108	RT_ELEMENTS(s_aSubTestsLzCnt16), s_aSubTestsLzCnt16 },
2109	{ BS3_CMN_NM(bs3CpuInstr2_f2_lzcnt_EAX_EBX_ud2), false, 5 + (ARCH_BITS == 16), 32, X86_EFL_ZF \| X86_EFL_CF,
2110	RT_ELEMENTS(s_aSubTestsLzCnt32), s_aSubTestsLzCnt32 },
2111	{ BS3_CMN_NM(bs3CpuInstr2_f2_lzcnt_EAX_FSxBX_ud2),true, 6 + (ARCH_BITS == 16), 32, X86_EFL_ZF \| X86_EFL_CF,
2112	RT_ELEMENTS(s_aSubTestsLzCnt32), s_aSubTestsLzCnt32 },
2113	#if ARCH_BITS == 64
2114	{ BS3_CMN_NM(bs3CpuInstr2_f2_lzcnt_RAX_RBX_ud2), false, 6, 64, X86_EFL_ZF \| X86_EFL_CF,
2115	RT_ELEMENTS(s_aSubTestsLzCnt64), s_aSubTestsLzCnt64 },
2116	{ BS3_CMN_NM(bs3CpuInstr2_f2_lzcnt_RAX_FSxBX_ud2),true, 7, 64, X86_EFL_ZF \| X86_EFL_CF,
2117	RT_ELEMENTS(s_aSubTestsLzCnt64), s_aSubTestsLzCnt64 },
2118	#endif
2119	};
2120
2121	uint32_t uExtFeatEcx = 0;
2122	if (g_uBs3CpuDetected & BS3CPU_F_CPUID_EXT_LEAVES)
2123	ASMCpuIdExSlow(UINT32_C(0x80000001), 0, 0, 0, NULL, NULL, &uExtFeatEcx, NULL);
2124	if (!(uExtFeatEcx & X86_CPUID_AMD_FEATURE_ECX_ABM))
2125	{
2126	unsigned i = RT_ELEMENTS(s_aTests);
2127	while (i-- > 0)
2128	if (s_aTests[i].fEflCheck & X86_EFL_CF)
2129	{
2130	s_aTests[i].fEflCheck = X86_EFL_ZF;
2131	switch (s_aTests[i].cOpBits)
2132	{
2133	case 16:
2134	s_aTests[i].cSubTests = RT_ELEMENTS(s_aSubTestsBsr16);
2135	s_aTests[i].paSubTests = s_aSubTestsBsr16;
2136	break;
2137	case 32:
2138	s_aTests[i].cSubTests = RT_ELEMENTS(s_aSubTestsBsr32);
2139	s_aTests[i].paSubTests = s_aSubTestsBsr32;
2140	break;
2141	#if ARCH_BITS == 64
2142	case 64:
2143	s_aTests[i].cSubTests = RT_ELEMENTS(s_aSubTestsBsr64);
2144	s_aTests[i].paSubTests = s_aSubTestsBsr64;
2145	break;
2146	#endif
2147	}
2148	}
2149	Bs3TestPrintf("lzcnt not supported\n");
2150	}
2151
2152	return bs3CpuInstr2_BitScan(bMode, s_aTests, RT_ELEMENTS(s_aTests));
2153	}
2154
2155
2156	/**
2157	* RORX
2158	*/
2159	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_rorx)(uint8_t bMode)
2160	{
2161	static const struct
2162	{
2163	FPFNBS3FAR pfnWorker;
2164	bool fMemSrc;
2165	bool fOkay;
2166	RTCCUINTXREG uIn;
2167	RTCCUINTXREG uOut;
2168	} s_aTests[] =
2169	{
2170	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
2171	{ BS3_CMN_NM(bs3CpuInstr2_rorx_RBX_RDX_2_icebp), false, true, // #0
2172	0, /* -> */ 0 },
2173	{ BS3_CMN_NM(bs3CpuInstr2_rorx_RBX_RDX_2_icebp), false, true, // #1
2174	~(RTCCUINTXREG)2, /* -> */ ~(RTCCUINTXREG)0 >> 1 },
2175	{ BS3_CMN_NM(bs3CpuInstr2_rorx_RBX_DSxDI_68_icebp), true, true, // #2
2176	0, /* -> */ 0 },
2177	{ BS3_CMN_NM(bs3CpuInstr2_rorx_RBX_DSxDI_68_icebp), true, true, // #3
2178	~(RTCCUINTXREG)2, /* -> / (RTCCUINTXREG_MAX >> 4) \| (~(RTCCUINTXREG)2 << (sizeof(RTCCUINTXREG) 8 - 4)) },
2179
2180	/* 32 bits register width: */
2181	{ BS3_CMN_NM(bs3CpuInstr2_rorx_EBX_EDX_2_icebp), false, true, // #4
2182	0, /* -> */ 0 },
2183	{ BS3_CMN_NM(bs3CpuInstr2_rorx_EBX_EDX_2_icebp), false, true, // #5
2184	~(RTCCUINTXREG)2, /* -> */ (RTCCUINTXREG)(~(uint32_t)0 >> 1) },
2185	{ BS3_CMN_NM(bs3CpuInstr2_rorx_EBX_DSxDI_36_icebp), true, true, // #6
2186	0, /* -> */ 0 },
2187	{ BS3_CMN_NM(bs3CpuInstr2_rorx_EBX_DSxDI_36_icebp), true, true, // #7
2188	~(RTCCUINTXREG)2, /* -> */ (RTCCUINTXREG)UINT32_C(0xdfffffff) },
2189
2190	/* encoding tests: */
2191	{ BS3_CMN_NM(bs3CpuInstr2_rorx_EBX_EDX_2_icebp_L1), false, false, // #8
2192	RTCCUINTXREG_MAX, /* -> */ 0 },
2193	{ BS3_CMN_NM(bs3CpuInstr2_rorx_EBX_EDX_2_icebp_V1), false, false, // #9
2194	RTCCUINTXREG_MAX, /* -> */ 0 },
2195	{ BS3_CMN_NM(bs3CpuInstr2_rorx_EBX_EDX_2_icebp_V15), false, false, // #10
2196	RTCCUINTXREG_MAX, /* -> */ 0 },
2197	# if ARCH_BITS == 64 /* The VEX.X=0 encoding mean LES instruction in 32-bit and 16-bit mode. */
2198	{ BS3_CMN_NM(bs3CpuInstr2_rorx_EBX_EDX_2_icebp_X1), false, true, // #11
2199	UINT32_C(0xf1e2d3c5), /* -> */ (RTCCUINTXREG)UINT32_C(0x7c78b4f1) },
2200	# endif
2201	};
2202
2203	BS3REGCTX Ctx;
2204	BS3TRAPFRAME TrapFrame;
2205	unsigned i, j;
2206	uint32_t uStdExtFeatEbx = 0;
2207	bool fSupportsRorX;
2208
2209	if (g_uBs3CpuDetected & BS3CPU_F_CPUID)
2210	ASMCpuIdExSlow(7, 0, 0, 0, NULL, &uStdExtFeatEbx, NULL, NULL);
2211	fSupportsRorX = RT_BOOL(uStdExtFeatEbx & X86_CPUID_STEXT_FEATURE_EBX_BMI2);
2212
2213	/* Ensure the structures are allocated before we sample the stack pointer. */
2214	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
2215	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
2216
2217	/*
2218	* Create test context.
2219	*/
2220	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
2221
2222	/*
2223	* Do the tests twice, first with all flags set, then once again with
2224	* flags cleared. The flags are not supposed to be touched at all.
2225	*/
2226	Ctx.rflags.u16 \|= X86_EFL_STATUS_BITS;
2227	for (j = 0; j < 2; j++)
2228	{
2229	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
2230	{
2231	bool const fOkay = !BS3_MODE_IS_RM_OR_V86(bMode) && s_aTests[i].fOkay && fSupportsRorX;
2232	uint8_t const bExpectXcpt = fOkay ? X86_XCPT_DB : X86_XCPT_UD;
2233	uint64_t uExpectRbx, uExpectRip;
2234	RTCCUINTXREG uMemSrc, uMemSrcExpect;
2235	Ctx.rbx.uCcXReg = RTCCUINTXREG_MAX * 1019;
2236	if (!s_aTests[i].fMemSrc)
2237	{
2238	Ctx.rdx.uCcXReg = s_aTests[i].uIn;
2239	uMemSrcExpect = uMemSrc = ~s_aTests[i].uIn;
2240	}
2241	else
2242	{
2243	Ctx.rdx.uCcXReg = ~s_aTests[i].uIn;
2244	uMemSrcExpect = uMemSrc = s_aTests[i].uIn;
2245	Bs3RegCtxSetGrpDsFromCurPtr(&Ctx, &Ctx.rdi, &uMemSrc);
2246	}
2247	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, s_aTests[i].pfnWorker);
2248	uExpectRbx = fOkay ? s_aTests[i].uOut : Ctx.rbx.u;
2249	uExpectRip = Ctx.rip.u + (fOkay ? 6 + 1 : 0);
2250	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
2251
2252	if ( TrapFrame.bXcpt != bExpectXcpt
2253	\|\| TrapFrame.Ctx.rip.u != uExpectRip
2254	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
2255	\|\| TrapFrame.Ctx.rbx.u != uExpectRbx
2256	/* check that nothing else really changed: */
2257	\|\| (TrapFrame.Ctx.rflags.u16 & X86_EFL_STATUS_BITS) != (Ctx.rflags.u16 & X86_EFL_STATUS_BITS)
2258	\|\| TrapFrame.Ctx.rax.u != Ctx.rax.u
2259	\|\| TrapFrame.Ctx.rcx.u != Ctx.rcx.u
2260	\|\| TrapFrame.Ctx.rsp.u != Ctx.rsp.u
2261	\|\| TrapFrame.Ctx.rbp.u != Ctx.rbp.u
2262	\|\| TrapFrame.Ctx.rsi.u != Ctx.rsi.u
2263	\|\| TrapFrame.Ctx.rdi.u != Ctx.rdi.u
2264	\|\| uMemSrc != uMemSrcExpect
2265	)
2266	{
2267	Bs3TestFailedF("test #%i failed: input %#" RTCCUINTXREG_XFMT, i, s_aTests[i].uIn);
2268	if (TrapFrame.bXcpt != bExpectXcpt)
2269	Bs3TestFailedF("Expected bXcpt = %#x, got %#x", bExpectXcpt, TrapFrame.bXcpt);
2270	if (TrapFrame.Ctx.rip.u != uExpectRip)
2271	Bs3TestFailedF("Expected RIP = %#06RX64, got %#06RX64", uExpectRip, TrapFrame.Ctx.rip.u);
2272	if (TrapFrame.Ctx.rdx.u != Ctx.rdx.u)
2273	Bs3TestFailedF("Expected RDX = %#06RX64, got %#06RX64 (src)", Ctx.rdx.u, TrapFrame.Ctx.rdx.u);
2274	if (TrapFrame.Ctx.rbx.u != uExpectRbx)
2275	Bs3TestFailedF("Expected RBX = %#06RX64, got %#06RX64 (dst)", uExpectRbx, TrapFrame.Ctx.rbx.u);
2276
2277	if ((TrapFrame.Ctx.rflags.u16 & X86_EFL_STATUS_BITS) != (Ctx.rflags.u16 & X86_EFL_STATUS_BITS))
2278	Bs3TestFailedF("Expected EFLAGS = %#06RX64, got %#06RX64",
2279	Ctx.rflags.u16 & X86_EFL_STATUS_BITS, TrapFrame.Ctx.rflags.u16 & X86_EFL_STATUS_BITS);
2280	if (TrapFrame.Ctx.rax.u != Ctx.rax.u)
2281	Bs3TestFailedF("Expected RAX = %#06RX64, got %#06RX64", Ctx.rax.u, TrapFrame.Ctx.rax.u);
2282	if (TrapFrame.Ctx.rcx.u != Ctx.rcx.u)
2283	Bs3TestFailedF("Expected RCX = %#06RX64, got %#06RX64", Ctx.rcx.u, TrapFrame.Ctx.rcx.u);
2284	if (TrapFrame.Ctx.rsp.u != Ctx.rsp.u)
2285	Bs3TestFailedF("Expected RSP = %#06RX64, got %#06RX64", Ctx.rsp.u, TrapFrame.Ctx.rsp.u);
2286	if (TrapFrame.Ctx.rbp.u != Ctx.rbp.u)
2287	Bs3TestFailedF("Expected RBP = %#06RX64, got %#06RX64", Ctx.rbp.u, TrapFrame.Ctx.rbp.u);
2288	if (TrapFrame.Ctx.rsi.u != Ctx.rsi.u)
2289	Bs3TestFailedF("Expected RSI = %#06RX64, got %#06RX64", Ctx.rsi.u, TrapFrame.Ctx.rsi.u);
2290	if (TrapFrame.Ctx.rdi.u != Ctx.rdi.u)
2291	Bs3TestFailedF("Expected RDI = %#06RX64, got %#06RX64", Ctx.rdi.u, TrapFrame.Ctx.rdi.u);
2292	if (uMemSrc != uMemSrcExpect)
2293	Bs3TestFailedF("Expected uMemSrc = %#06RX64, got %#06RX64", (uint64_t)uMemSrcExpect, (uint64_t)uMemSrc);
2294	}
2295	}
2296	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
2297	}
2298
2299	return 0;
2300	}
2301
2302
2303	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_andn)(uint8_t bMode)
2304	{
2305	#define ANDN_CHECK_EFLAGS (uint16_t)(X86_EFL_CF \| X86_EFL_ZF \| X86_EFL_OF \| X86_EFL_SF)
2306	#define ANDN_IGNORE_EFLAGS (uint16_t)(X86_EFL_AF \| X86_EFL_PF) /* undefined, ignoring for now */
2307	static const struct
2308	{
2309	FPFNBS3FAR pfnWorker;
2310	bool fMemSrc;
2311	uint8_t cbInstr;
2312	RTCCUINTXREG uSrc1;
2313	RTCCUINTXREG uSrc2;
2314	RTCCUINTXREG uOut;
2315	uint16_t fEFlags;
2316	} s_aTests[] =
2317	{
2318	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
2319	{ BS3_CMN_NM(bs3CpuInstr2_andn_RAX_RCX_RBX_icebp), false, 5, // #0
2320	0, 0, /* -> */ 0, X86_EFL_ZF },
2321	{ BS3_CMN_NM(bs3CpuInstr2_andn_RAX_RCX_RBX_icebp), false, 5, // #1
2322	2, ~(RTCCUINTXREG)3, /* -> */ ~(RTCCUINTXREG)3, X86_EFL_SF },
2323	{ BS3_CMN_NM(bs3CpuInstr2_andn_RAX_RCX_FSxBX_icebp), true, 6, // #2
2324	0, 0, /* -> */ 0, X86_EFL_ZF },
2325	{ BS3_CMN_NM(bs3CpuInstr2_andn_RAX_RCX_FSxBX_icebp), true, 6, // #3
2326	2, ~(RTCCUINTXREG)3, /* -> */ ~(RTCCUINTXREG)3, X86_EFL_SF },
2327
2328	/* 32-bit register width */
2329	{ BS3_CMN_NM(bs3CpuInstr2_andn_EAX_ECX_EBX_icebp), false, 5, // #4
2330	0, 0, /* -> */ 0, X86_EFL_ZF },
2331	{ BS3_CMN_NM(bs3CpuInstr2_andn_EAX_ECX_EBX_icebp), false, 5, // #5
2332	2, ~(RTCCUINTXREG)7, /* -> */ ~(uint32_t)7, X86_EFL_SF },
2333	{ BS3_CMN_NM(bs3CpuInstr2_andn_EAX_ECX_FSxBX_icebp), true, 6, // #6
2334	0, 0, /* -> */ 0, X86_EFL_ZF },
2335	{ BS3_CMN_NM(bs3CpuInstr2_andn_EAX_ECX_FSxBX_icebp), true, 6, // #7
2336	2, ~(RTCCUINTXREG)7, /* -> */ ~(uint32_t)7, X86_EFL_SF },
2337
2338	};
2339
2340	BS3REGCTX Ctx;
2341	BS3TRAPFRAME TrapFrame;
2342	unsigned i, j;
2343	uint32_t uStdExtFeatEbx = 0;
2344	bool fSupportsAndN;
2345
2346	if (g_uBs3CpuDetected & BS3CPU_F_CPUID)
2347	ASMCpuIdExSlow(7, 0, 0, 0, NULL, &uStdExtFeatEbx, NULL, NULL);
2348	fSupportsAndN = RT_BOOL(uStdExtFeatEbx & X86_CPUID_STEXT_FEATURE_EBX_BMI1);
2349
2350	/* Ensure the structures are allocated before we sample the stack pointer. */
2351	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
2352	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
2353
2354	/*
2355	* Create test context.
2356	*/
2357	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
2358
2359	/*
2360	* Do the tests twice, first with all flags set, then once again with
2361	* flags cleared. The flags are not supposed to be touched at all.
2362	*/
2363	Ctx.rflags.u16 \|= X86_EFL_STATUS_BITS;
2364	for (j = 0; j < 2; j++)
2365	{
2366	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
2367	{
2368	bool const fOkay = !BS3_MODE_IS_RM_OR_V86(bMode) && fSupportsAndN;
2369	uint8_t const bExpectXcpt = fOkay ? X86_XCPT_DB : X86_XCPT_UD;
2370	uint64_t uExpectRax, uExpectRip;
2371	RTCCUINTXREG uMemSrc2, uMemSrc2Expect;
2372
2373	Ctx.rax.uCcXReg = RTCCUINTXREG_MAX * 1019;
2374	Ctx.rcx.uCcXReg = s_aTests[i].uSrc1;
2375	if (!s_aTests[i].fMemSrc)
2376	{
2377	Ctx.rbx.uCcXReg = s_aTests[i].uSrc2;
2378	uMemSrc2Expect = uMemSrc2 = ~s_aTests[i].uSrc2;
2379	}
2380	else
2381	{
2382	uMemSrc2Expect = uMemSrc2 = s_aTests[i].uSrc2;
2383	Bs3RegCtxSetGrpSegFromCurPtr(&Ctx, &Ctx.rbx, &Ctx.fs, &uMemSrc2);
2384	}
2385	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, s_aTests[i].pfnWorker);
2386	uExpectRax = fOkay ? s_aTests[i].uOut : Ctx.rax.u;
2387	uExpectRip = Ctx.rip.u + (fOkay ? s_aTests[i].cbInstr + 1 : 0);
2388	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
2389
2390	if ( TrapFrame.bXcpt != bExpectXcpt
2391	\|\| TrapFrame.Ctx.rip.u != uExpectRip
2392	\|\| TrapFrame.Ctx.rcx.u != Ctx.rcx.u
2393	\|\| TrapFrame.Ctx.rbx.u != Ctx.rbx.u
2394	\|\| TrapFrame.Ctx.rax.u != uExpectRax
2395	/* check that nothing else really changed: */
2396	\|\| (TrapFrame.Ctx.rflags.u16 & ANDN_CHECK_EFLAGS)
2397	!= ((fOkay ? s_aTests[i].fEFlags : Ctx.rflags.u16) & ANDN_CHECK_EFLAGS)
2398	\|\| (TrapFrame.Ctx.rflags.u16 & ~(ANDN_CHECK_EFLAGS \| ANDN_IGNORE_EFLAGS) & X86_EFL_STATUS_BITS)
2399	!= (Ctx.rflags.u16 & ~(ANDN_CHECK_EFLAGS \| ANDN_IGNORE_EFLAGS) & X86_EFL_STATUS_BITS)
2400	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
2401	\|\| TrapFrame.Ctx.rsp.u != Ctx.rsp.u
2402	\|\| TrapFrame.Ctx.rbp.u != Ctx.rbp.u
2403	\|\| TrapFrame.Ctx.rsi.u != Ctx.rsi.u
2404	\|\| TrapFrame.Ctx.rdi.u != Ctx.rdi.u
2405	\|\| uMemSrc2 != uMemSrc2Expect
2406	)
2407	{
2408	Bs3TestFailedF("test #%i failed: input %#" RTCCUINTXREG_XFMT ", %#" RTCCUINTXREG_XFMT, i, s_aTests[i].uSrc1, s_aTests[i].uSrc2);
2409	if (TrapFrame.bXcpt != bExpectXcpt)
2410	Bs3TestFailedF("Expected bXcpt = %#x, got %#x", bExpectXcpt, TrapFrame.bXcpt);
2411	if (TrapFrame.Ctx.rip.u != uExpectRip)
2412	Bs3TestFailedF("Expected RIP = %#06RX64, got %#06RX64", uExpectRip, TrapFrame.Ctx.rip.u);
2413	if (TrapFrame.Ctx.rax.u != uExpectRax)
2414	Bs3TestFailedF("Expected RAX = %#06RX64, got %#06RX64", uExpectRax, TrapFrame.Ctx.rax.u);
2415	if (TrapFrame.Ctx.rcx.u != Ctx.rcx.u)
2416	Bs3TestFailedF("Expected RCX = %#06RX64, got %#06RX64", Ctx.rcx.u, TrapFrame.Ctx.rcx.u);
2417	if (TrapFrame.Ctx.rbx.u != Ctx.rbx.u)
2418	Bs3TestFailedF("Expected RBX = %#06RX64, got %#06RX64 (dst)", Ctx.rbx.u, TrapFrame.Ctx.rbx.u);
2419	if ( (TrapFrame.Ctx.rflags.u16 & ANDN_CHECK_EFLAGS)
2420	!= ((fOkay ? s_aTests[i].fEFlags : Ctx.rflags.u16) & ANDN_CHECK_EFLAGS))
2421	Bs3TestFailedF("Expected EFLAGS = %#06RX32, got %#06RX32 (output)",
2422	(fOkay ? s_aTests[i].fEFlags : Ctx.rflags.u16) & ANDN_CHECK_EFLAGS, TrapFrame.Ctx.rflags.u16 & ANDN_CHECK_EFLAGS);
2423	if ( (TrapFrame.Ctx.rflags.u16 & ~(ANDN_CHECK_EFLAGS \| ANDN_IGNORE_EFLAGS) & X86_EFL_STATUS_BITS)
2424	!= (Ctx.rflags.u16 & ~(ANDN_CHECK_EFLAGS \| ANDN_IGNORE_EFLAGS) & X86_EFL_STATUS_BITS))
2425	Bs3TestFailedF("Expected EFLAGS = %#06RX32, got %#06RX32 (immutable)",
2426	Ctx.rflags.u16 & ~(ANDN_CHECK_EFLAGS \| ANDN_IGNORE_EFLAGS) & X86_EFL_STATUS_BITS,
2427	TrapFrame.Ctx.rflags.u16 & ~(ANDN_CHECK_EFLAGS \| ANDN_IGNORE_EFLAGS) & X86_EFL_STATUS_BITS);
2428
2429	if (TrapFrame.Ctx.rdx.u != Ctx.rdx.u)
2430	Bs3TestFailedF("Expected RDX = %#06RX64, got %#06RX64 (src)", Ctx.rdx.u, TrapFrame.Ctx.rdx.u);
2431	if (TrapFrame.Ctx.rsp.u != Ctx.rsp.u)
2432	Bs3TestFailedF("Expected RSP = %#06RX64, got %#06RX64", Ctx.rsp.u, TrapFrame.Ctx.rsp.u);
2433	if (TrapFrame.Ctx.rbp.u != Ctx.rbp.u)
2434	Bs3TestFailedF("Expected RBP = %#06RX64, got %#06RX64", Ctx.rbp.u, TrapFrame.Ctx.rbp.u);
2435	if (TrapFrame.Ctx.rsi.u != Ctx.rsi.u)
2436	Bs3TestFailedF("Expected RSI = %#06RX64, got %#06RX64", Ctx.rsi.u, TrapFrame.Ctx.rsi.u);
2437	if (TrapFrame.Ctx.rdi.u != Ctx.rdi.u)
2438	Bs3TestFailedF("Expected RDI = %#06RX64, got %#06RX64", Ctx.rdi.u, TrapFrame.Ctx.rdi.u);
2439	if (uMemSrc2 != uMemSrc2Expect)
2440	Bs3TestFailedF("Expected uMemSrc2 = %#06RX64, got %#06RX64", (uint64_t)uMemSrc2Expect, (uint64_t)uMemSrc2);
2441	}
2442	}
2443	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
2444	}
2445
2446	return 0;
2447	}
2448
2449	/*
2450	* For testing BEXTR, SHLX SARX & SHRX.
2451	*/
2452	typedef struct BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T
2453	{
2454	RTCCUINTXREG uSrc1;
2455	RTCCUINTXREG uSrc2;
2456	RTCCUINTXREG uOut;
2457	uint16_t fEflOut;
2458	} BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T;
2459
2460	typedef struct BS3CPUINSTR2_TEST_Gy_Ey_By_T
2461	{
2462	FPFNBS3FAR pfnWorker;
2463	bool fMemSrc;
2464	uint8_t cbInstr;
2465	uint8_t cSubTests;
2466	BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const *paSubTests;
2467	} BS3CPUINSTR2_TEST_Gy_Ey_By_T;
2468
2469	static uint8_t bs3CpuInstr2_Common_Gy_Ey_By(uint8_t bMode, BS3CPUINSTR2_TEST_Gy_Ey_By_T const *paTests, unsigned cTests,
2470	uint32_t fStdExtFeatEbx, uint16_t fEflCheck, uint16_t fEflIgnore)
2471	{
2472	BS3REGCTX Ctx;
2473	BS3TRAPFRAME TrapFrame;
2474	unsigned i, j, k;
2475	uint32_t uStdExtFeatEbx = 0;
2476	bool fSupportsInstr;
2477
2478	fEflCheck &= ~fEflIgnore;
2479
2480	if (g_uBs3CpuDetected & BS3CPU_F_CPUID)
2481	ASMCpuIdExSlow(7, 0, 0, 0, NULL, &uStdExtFeatEbx, NULL, NULL);
2482	fSupportsInstr = RT_BOOL(uStdExtFeatEbx & fStdExtFeatEbx);
2483
2484	/* Ensure the structures are allocated before we sample the stack pointer. */
2485	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
2486	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
2487
2488	/*
2489	* Create test context.
2490	*/
2491	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
2492
2493	/*
2494	* Do the tests twice, first with all flags set, then once again with
2495	* flags cleared. The flags are not supposed to be touched at all.
2496	*/
2497	Ctx.rflags.u16 \|= X86_EFL_STATUS_BITS;
2498	for (j = 0; j < 2; j++)
2499	{
2500	for (i = 0; i < cTests; i++)
2501	{
2502	for (k = 0; k < paTests[i].cSubTests; k++)
2503	{
2504	bool const fOkay = !BS3_MODE_IS_RM_OR_V86(bMode) && fSupportsInstr;
2505	uint8_t const bExpectXcpt = fOkay ? X86_XCPT_DB : X86_XCPT_UD;
2506	uint64_t uExpectRax, uExpectRip;
2507	RTCCUINTXREG uMemSrc1, uMemSrc1Expect;
2508
2509	Ctx.rax.uCcXReg = RTCCUINTXREG_MAX * 1019;
2510	Ctx.rcx.uCcXReg = paTests[i].paSubTests[k].uSrc2;
2511	if (!paTests[i].fMemSrc)
2512	{
2513	Ctx.rbx.uCcXReg = paTests[i].paSubTests[k].uSrc1;
2514	uMemSrc1Expect = uMemSrc1 = ~paTests[i].paSubTests[k].uSrc1;
2515	}
2516	else
2517	{
2518	uMemSrc1Expect = uMemSrc1 = paTests[i].paSubTests[k].uSrc1;
2519	Bs3RegCtxSetGrpSegFromCurPtr(&Ctx, &Ctx.rbx, &Ctx.fs, &uMemSrc1);
2520	}
2521	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, paTests[i].pfnWorker);
2522	uExpectRax = fOkay ? paTests[i].paSubTests[k].uOut : Ctx.rax.u;
2523	uExpectRip = Ctx.rip.u + (fOkay ? paTests[i].cbInstr + 1 : 0);
2524	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
2525
2526	if ( TrapFrame.bXcpt != bExpectXcpt
2527	\|\| TrapFrame.Ctx.rip.u != uExpectRip
2528	\|\| TrapFrame.Ctx.rcx.u != Ctx.rcx.u
2529	\|\| TrapFrame.Ctx.rbx.u != Ctx.rbx.u
2530	\|\| TrapFrame.Ctx.rax.u != uExpectRax
2531	/* check that nothing else really changed: */
2532	\|\| (TrapFrame.Ctx.rflags.u16 & fEflCheck)
2533	!= ((fOkay ? paTests[i].paSubTests[k].fEflOut : Ctx.rflags.u16) & fEflCheck)
2534	\|\| (TrapFrame.Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS)
2535	!= (Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS)
2536	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
2537	\|\| TrapFrame.Ctx.rsp.u != Ctx.rsp.u
2538	\|\| TrapFrame.Ctx.rbp.u != Ctx.rbp.u
2539	\|\| TrapFrame.Ctx.rsi.u != Ctx.rsi.u
2540	\|\| TrapFrame.Ctx.rdi.u != Ctx.rdi.u
2541	\|\| uMemSrc1 != uMemSrc1Expect
2542	)
2543	{
2544	Bs3TestFailedF("test #%i/%i failed: input %#" RTCCUINTXREG_XFMT ", %#" RTCCUINTXREG_XFMT,
2545	i, k, paTests[i].paSubTests[k].uSrc1, paTests[i].paSubTests[k].uSrc2);
2546	if (TrapFrame.bXcpt != bExpectXcpt)
2547	Bs3TestFailedF("Expected bXcpt = %#x, got %#x", bExpectXcpt, TrapFrame.bXcpt);
2548	if (TrapFrame.Ctx.rip.u != uExpectRip)
2549	Bs3TestFailedF("Expected RIP = %#06RX64, got %#06RX64", uExpectRip, TrapFrame.Ctx.rip.u);
2550	if (TrapFrame.Ctx.rax.u != uExpectRax)
2551	Bs3TestFailedF("Expected RAX = %#06RX64, got %#06RX64", uExpectRax, TrapFrame.Ctx.rax.u);
2552	if (TrapFrame.Ctx.rcx.u != Ctx.rcx.u)
2553	Bs3TestFailedF("Expected RCX = %#06RX64, got %#06RX64", Ctx.rcx.u, TrapFrame.Ctx.rcx.u);
2554	if (TrapFrame.Ctx.rbx.u != Ctx.rbx.u)
2555	Bs3TestFailedF("Expected RBX = %#06RX64, got %#06RX64", Ctx.rbx.u, TrapFrame.Ctx.rbx.u);
2556	if ( (TrapFrame.Ctx.rflags.u16 & fEflCheck)
2557	!= ((fOkay ? paTests[i].paSubTests[k].fEflOut : Ctx.rflags.u16) & fEflCheck))
2558	Bs3TestFailedF("Expected EFLAGS = %#06RX32, got %#06RX32 (output)",
2559	(fOkay ? paTests[i].paSubTests[k].fEflOut : Ctx.rflags.u16) & fEflCheck,
2560	TrapFrame.Ctx.rflags.u16 & fEflCheck);
2561	if ( (TrapFrame.Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS)
2562	!= (Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS))
2563	Bs3TestFailedF("Expected EFLAGS = %#06RX32, got %#06RX32 (immutable)",
2564	Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS,
2565	TrapFrame.Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS);
2566
2567	if (TrapFrame.Ctx.rdx.u != Ctx.rdx.u)
2568	Bs3TestFailedF("Expected RDX = %#06RX64, got %#06RX64", Ctx.rdx.u, TrapFrame.Ctx.rdx.u);
2569	if (TrapFrame.Ctx.rsp.u != Ctx.rsp.u)
2570	Bs3TestFailedF("Expected RSP = %#06RX64, got %#06RX64", Ctx.rsp.u, TrapFrame.Ctx.rsp.u);
2571	if (TrapFrame.Ctx.rbp.u != Ctx.rbp.u)
2572	Bs3TestFailedF("Expected RBP = %#06RX64, got %#06RX64", Ctx.rbp.u, TrapFrame.Ctx.rbp.u);
2573	if (TrapFrame.Ctx.rsi.u != Ctx.rsi.u)
2574	Bs3TestFailedF("Expected RSI = %#06RX64, got %#06RX64", Ctx.rsi.u, TrapFrame.Ctx.rsi.u);
2575	if (TrapFrame.Ctx.rdi.u != Ctx.rdi.u)
2576	Bs3TestFailedF("Expected RDI = %#06RX64, got %#06RX64", Ctx.rdi.u, TrapFrame.Ctx.rdi.u);
2577	if (uMemSrc1 != uMemSrc1Expect)
2578	Bs3TestFailedF("Expected uMemSrc1 = %#06RX64, got %#06RX64", (uint64_t)uMemSrc1Expect, (uint64_t)uMemSrc1);
2579	}
2580	}
2581	}
2582	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
2583	}
2584
2585	return 0;
2586	}
2587
2588
2589	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_bextr)(uint8_t bMode)
2590	{
2591	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
2592	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests64[] =
2593	{
2594	{ 0, RT_MAKE_U16(0, 0), /* -> */ 0, X86_EFL_ZF },
2595	{ 0, RT_MAKE_U16(16, 33), /* -> */ 0, X86_EFL_ZF },
2596	{ ~(RTCCUINTXREG)7, RT_MAKE_U16(2, 4), /* -> */ 0xe, 0},
2597	{ ~(RTCCUINTXREG)7, RT_MAKE_U16(40, 8), /* -> */ ARCH_BITS == 64 ? 0xff : 0x00, ARCH_BITS == 64 ? 0 : X86_EFL_ZF },
2598	};
2599
2600	/* 32-bit register width */
2601	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests32[] =
2602	{
2603	{ 0, RT_MAKE_U16(0, 0), /* -> */ 0, X86_EFL_ZF },
2604	{ 0, RT_MAKE_U16(16, 18), /* -> */ 0, X86_EFL_ZF },
2605	{ ~(RTCCUINTXREG)7, RT_MAKE_U16(2, 4), /* -> */ 0xe, 0 },
2606	{ ~(RTCCUINTXREG)7, RT_MAKE_U16(24, 8), /* -> */ 0xff, 0 },
2607	{ ~(RTCCUINTXREG)7, RT_MAKE_U16(31, 9), /* -> */ 1, 0 },
2608	{ ~(RTCCUINTXREG)7, RT_MAKE_U16(42, 8), /* -> */ 0, X86_EFL_ZF },
2609	};
2610
2611	static BS3CPUINSTR2_TEST_Gy_Ey_By_T const s_aTests[] =
2612	{
2613	{ BS3_CMN_NM(bs3CpuInstr2_bextr_RAX_RBX_RCX_icebp), false, 5, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2614	{ BS3_CMN_NM(bs3CpuInstr2_bextr_RAX_FSxBX_RCX_icebp), true, 6, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2615	{ BS3_CMN_NM(bs3CpuInstr2_bextr_EAX_EBX_ECX_icebp), false, 5, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2616	{ BS3_CMN_NM(bs3CpuInstr2_bextr_EAX_FSxBX_ECX_icebp), true, 6, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2617	};
2618	return bs3CpuInstr2_Common_Gy_Ey_By(bMode, s_aTests, RT_ELEMENTS(s_aTests), X86_CPUID_STEXT_FEATURE_EBX_BMI1,
2619	X86_EFL_STATUS_BITS, X86_EFL_AF \| X86_EFL_SF \| X86_EFL_PF);
2620	}
2621
2622
2623	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_bzhi)(uint8_t bMode)
2624	{
2625	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
2626	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests64[] =
2627	{
2628	{ 0, 0, /* -> */ 0, X86_EFL_ZF },
2629	{ 0, ~(RTCCUINTXREG)255, /* -> */ 0, X86_EFL_ZF },
2630	{ 0, 64, /* -> */ 0, X86_EFL_ZF \| X86_EFL_CF },
2631	{ ~(RTCCUINTXREG)0, 64, /* -> */ ~(RTCCUINTXREG)0, X86_EFL_CF \| X86_EFL_SF },
2632	{ ~(RTCCUINTXREG)0, 63,
2633	/* -> */ ARCH_BITS >= 64 ? ~(RTCCUINTXREG)0 >> 1 : ~(RTCCUINTXREG)0, ARCH_BITS >= 64 ? 0 : X86_EFL_CF \| X86_EFL_SF },
2634	{ ~(RTCCUINTXREG)0 << 31 \| UINT32_C(0x63849607), 24, /* -> */ UINT32_C(0x00849607), 0 },
2635	{ ~(RTCCUINTXREG)0 << 31 \| UINT32_C(0x63849607), 33,
2636	/* -> */ ARCH_BITS >= 64 ? UINT64_C(0x1e3849607) : UINT32_C(0xe3849607), ARCH_BITS >= 64 ? 0 : X86_EFL_CF \| X86_EFL_SF },
2637	};
2638
2639	/* 32-bit register width */
2640	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests32[] =
2641	{
2642	{ 0, 0, /* -> */ 0, X86_EFL_ZF },
2643	{ 0, ~(RTCCUINTXREG)255, /* -> */ 0, X86_EFL_ZF },
2644	{ 0, 32, /* -> */ 0, X86_EFL_ZF \| X86_EFL_CF },
2645	{ ~(RTCCUINTXREG)0, 32, /* -> */ UINT32_MAX, X86_EFL_CF \| X86_EFL_SF },
2646	{ ~(RTCCUINTXREG)0, 31, /* -> */ UINT32_MAX >> 1, 0 },
2647	{ UINT32_C(0x1230fd34), 15, /* -> */ UINT32_C(0x00007d34), 0 },
2648	};
2649
2650	static BS3CPUINSTR2_TEST_Gy_Ey_By_T const s_aTests[] =
2651	{
2652	{ BS3_CMN_NM(bs3CpuInstr2_bzhi_RAX_RBX_RCX_icebp), false, 5, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2653	{ BS3_CMN_NM(bs3CpuInstr2_bzhi_RAX_FSxBX_RCX_icebp), true, 6, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2654	{ BS3_CMN_NM(bs3CpuInstr2_bzhi_EAX_EBX_ECX_icebp), false, 5, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2655	{ BS3_CMN_NM(bs3CpuInstr2_bzhi_EAX_FSxBX_ECX_icebp), true, 6, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2656	};
2657	return bs3CpuInstr2_Common_Gy_Ey_By(bMode, s_aTests, RT_ELEMENTS(s_aTests), X86_CPUID_STEXT_FEATURE_EBX_BMI2,
2658	X86_EFL_STATUS_BITS, 0);
2659	}
2660
2661
2662	/** @note This is a Gy_By_Ey format instruction, so we're switching the two
2663	* source registers around when calling bs3CpuInstr2_Common_Gy_Ey_By.
2664	* Sorry for the confusion, but it saves some unnecessary code dup. */
2665	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_pdep)(uint8_t bMode)
2666	{
2667	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
2668	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests64[] =
2669	{ /* Mask (RBX/[FS:xBX]), source=RCX */
2670	{ 0, 0, /* -> */ 0, 0 },
2671	{ 0, ~(RTCCUINTXREG)0, /* -> */ 0, 0 },
2672	{ ~(RTCCUINTXREG)0, 0, /* -> */ 0, 0 },
2673	{ ~(RTCCUINTXREG)0, ~(RTCCUINTXREG)0, /* -> */ ~(RTCCUINTXREG)0, 0 },
2674	#if ARCH_BITS >= 64
2675	{ UINT64_C(0x3586049947589201), ~(RTCCUINTXREG)0, /* -> */ UINT64_C(0x3586049947589201), 0 },
2676	{ UINT64_C(0x3586049947589201), ~(RTCCUINTXREG)7, /* -> */ UINT64_C(0x3586049947588000), 0 },
2677	#endif
2678	{ UINT32_C(0x47589201), ~(RTCCUINTXREG)0, /* -> */ UINT32_C(0x47589201), 0 },
2679	{ UINT32_C(0x47589201), ~(RTCCUINTXREG)7, /* -> */ UINT32_C(0x47588000), 0 },
2680	};
2681
2682	/* 32-bit register width */
2683	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests32[] =
2684	{ /* Mask (EBX/[FS:xBX]), source=ECX */
2685	{ 0, 0, /* -> */ 0, 0 },
2686	{ 0, ~(RTCCUINTXREG)0, /* -> */ 0, 0 },
2687	{ ~(RTCCUINTXREG)0, 0, /* -> */ 0, 0 },
2688	{ ~(RTCCUINTXREG)0, ~(RTCCUINTXREG)0, /* -> */ UINT32_MAX, 0 },
2689	{ UINT32_C(0x01010101), ~(RTCCUINTXREG)0, /* -> */ UINT32_C(0x01010101), 0 },
2690	{ UINT32_C(0x01010101), ~(RTCCUINTXREG)3, /* -> */ UINT32_C(0x01010000), 0 },
2691	{ UINT32_C(0x47589201), ~(RTCCUINTXREG)0, /* -> */ UINT32_C(0x47589201), 0 },
2692	};
2693
2694	static BS3CPUINSTR2_TEST_Gy_Ey_By_T const s_aTests[] =
2695	{
2696	{ BS3_CMN_NM(bs3CpuInstr2_pdep_RAX_RCX_RBX_icebp), false, 5, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2697	{ BS3_CMN_NM(bs3CpuInstr2_pdep_RAX_RCX_FSxBX_icebp), true, 6, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2698	{ BS3_CMN_NM(bs3CpuInstr2_pdep_EAX_ECX_EBX_icebp), false, 5, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2699	{ BS3_CMN_NM(bs3CpuInstr2_pdep_EAX_ECX_FSxBX_icebp), true, 6, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2700	};
2701	return bs3CpuInstr2_Common_Gy_Ey_By(bMode, s_aTests, RT_ELEMENTS(s_aTests), X86_CPUID_STEXT_FEATURE_EBX_BMI2, 0, 0);
2702	}
2703
2704
2705	/** @note Same note as for bs3CpuInstr2_pdep */
2706	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_pext)(uint8_t bMode)
2707	{
2708	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
2709	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests64[] =
2710	{ /* Mask (RBX/[FS:xBX]), source=RCX */
2711	{ 0, 0, /* -> */ 0, 0 },
2712	{ 0, ~(RTCCUINTXREG)0, /* -> */ 0, 0 },
2713	{ ~(RTCCUINTXREG)0, 0, /* -> */ 0, 0 },
2714	{ ~(RTCCUINTXREG)0, ~(RTCCUINTXREG)0, /* -> */ ~(RTCCUINTXREG)0, 0 },
2715	#if ARCH_BITS >= 64
2716	{ UINT64_C(0x3586049947589201), ~(RTCCUINTXREG)0, /* -> */ UINT64_C(0x00000000007fffff), 0 },
2717	{ UINT64_C(0x3586049947589201), ~(RTCCUINTXREG)7, /* -> */ UINT64_C(0x00000000007ffffe), 0 },
2718	#endif
2719	{ UINT32_C(0x47589201), ~(RTCCUINTXREG)0, /* -> */ UINT32_C(0x000007ff), 0 },
2720	{ UINT32_C(0x47589201), ~(RTCCUINTXREG)7, /* -> */ UINT32_C(0x000007fe), 0 },
2721	};
2722
2723	/* 32-bit register width */
2724	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests32[] =
2725	{ /* Mask (EBX/[FS:xBX]), source=ECX */
2726	{ 0, 0, /* -> */ 0, 0 },
2727	{ 0, ~(RTCCUINTXREG)0, /* -> */ 0, 0 },
2728	{ ~(RTCCUINTXREG)0, 0, /* -> */ 0, 0 },
2729	{ ~(RTCCUINTXREG)0, ~(RTCCUINTXREG)0, /* -> */ UINT32_MAX, 0 },
2730	{ UINT32_C(0x01010101), ~(RTCCUINTXREG)0, /* -> */ UINT32_C(0x0000000f), 0 },
2731	{ UINT32_C(0x01010101), ~(RTCCUINTXREG)3, /* -> */ UINT32_C(0x0000000e), 0 },
2732	{ UINT32_C(0x47589201), ~(RTCCUINTXREG)0, /* -> */ UINT32_C(0x000007ff), 0 },
2733	{ UINT32_C(0x47589201), ~(RTCCUINTXREG)7, /* -> */ UINT32_C(0x000007fe), 0 },
2734	};
2735
2736	static BS3CPUINSTR2_TEST_Gy_Ey_By_T const s_aTests[] =
2737	{
2738	{ BS3_CMN_NM(bs3CpuInstr2_pext_RAX_RCX_RBX_icebp), false, 5, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2739	{ BS3_CMN_NM(bs3CpuInstr2_pext_RAX_RCX_FSxBX_icebp), true, 6, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2740	{ BS3_CMN_NM(bs3CpuInstr2_pext_EAX_ECX_EBX_icebp), false, 5, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2741	{ BS3_CMN_NM(bs3CpuInstr2_pext_EAX_ECX_FSxBX_icebp), true, 6, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2742	};
2743	return bs3CpuInstr2_Common_Gy_Ey_By(bMode, s_aTests, RT_ELEMENTS(s_aTests), X86_CPUID_STEXT_FEATURE_EBX_BMI2, 0, 0);
2744	}
2745
2746
2747	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_shlx)(uint8_t bMode)
2748	{
2749	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
2750	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests64[] =
2751	{
2752	{ 0, 0, /* -> */ 0, 0 },
2753	{ 0, ~(RTCCUINTXREG)3, /* -> */ 0, 0 },
2754	{ ~(RTCCUINTXREG)7, 8, /* -> */ ~(RTCCUINTXREG)0x7ff, 0},
2755	{ ~(RTCCUINTXREG)7, 40, /* -> */ ~(RTCCUINTXREG)7 << (ARCH_BITS == 64 ? 40 : 8), 0 },
2756	{ ~(RTCCUINTXREG)7, 72, /* -> */ ~(RTCCUINTXREG)7 << 8, 0 },
2757	};
2758
2759	/* 32-bit register width */
2760	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests32[] =
2761	{
2762	{ 0, 0, /* -> */ 0, 0 },
2763	{ 0, ~(RTCCUINTXREG)9, /* -> */ 0, 0 },
2764	{ ~(RTCCUINTXREG)7, 8, /* -> */ UINT32_C(0xfffff800), 0 },
2765	{ ~(RTCCUINTXREG)7, 8, /* -> */ UINT32_C(0xfffff800), 0 },
2766	};
2767
2768	static BS3CPUINSTR2_TEST_Gy_Ey_By_T const s_aTests[] =
2769	{
2770	{ BS3_CMN_NM(bs3CpuInstr2_shlx_RAX_RBX_RCX_icebp), false, 5, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2771	{ BS3_CMN_NM(bs3CpuInstr2_shlx_RAX_FSxBX_RCX_icebp), true, 6, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2772	{ BS3_CMN_NM(bs3CpuInstr2_shlx_EAX_EBX_ECX_icebp), false, 5, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2773	{ BS3_CMN_NM(bs3CpuInstr2_shlx_EAX_FSxBX_ECX_icebp), true, 6, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2774	};
2775	return bs3CpuInstr2_Common_Gy_Ey_By(bMode, s_aTests, RT_ELEMENTS(s_aTests), X86_CPUID_STEXT_FEATURE_EBX_BMI1,
2776	0, 0);
2777	}
2778
2779
2780	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_sarx)(uint8_t bMode)
2781	{
2782	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
2783	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests64[] =
2784	{
2785	{ 0, 0, /* -> */ 0, 0 },
2786	{ 0, ~(RTCCUINTXREG)3, /* -> */ 0, 0 },
2787	{ (RTCCUINTXREG)1 << (RTCCINTXREG_BITS - 1), RTCCINTXREG_BITS - 1, /* -> */ ~(RTCCUINTXREG)0, 0 },
2788	{ (RTCCUINTXREG)1 << (RTCCINTXREG_BITS - 1), RTCCINTXREG_BITS - 1 + 64, /* -> */ ~(RTCCUINTXREG)0, 0 },
2789	{ (RTCCUINTXREG)1 << (RTCCINTXREG_BITS - 2), RTCCINTXREG_BITS - 3, /* -> */ 2, 0 },
2790	{ (RTCCUINTXREG)1 << (RTCCINTXREG_BITS - 2), RTCCINTXREG_BITS - 3 + 64, /* -> */ 2, 0 },
2791	};
2792
2793	/* 32-bit register width */
2794	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests32[] =
2795	{
2796	{ 0, 0, /* -> */ 0, 0 },
2797	{ 0, ~(RTCCUINTXREG)9, /* -> */ 0, 0 },
2798	{ ~(RTCCUINTXREG)UINT32_C(0x7fffffff), 24, /* -> */ UINT32_C(0xffffff80), 0 },
2799	{ ~(RTCCUINTXREG)UINT32_C(0x7fffffff), 24+32, /* -> */ UINT32_C(0xffffff80), 0 },
2800	{ ~(RTCCUINTXREG)UINT32_C(0xbfffffff), 24, /* -> */ UINT32_C(0x40), 0 },
2801	{ ~(RTCCUINTXREG)UINT32_C(0xbfffffff), 24+32, /* -> */ UINT32_C(0x40), 0 },
2802	};
2803
2804	static BS3CPUINSTR2_TEST_Gy_Ey_By_T const s_aTests[] =
2805	{
2806	{ BS3_CMN_NM(bs3CpuInstr2_sarx_RAX_RBX_RCX_icebp), false, 5, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2807	{ BS3_CMN_NM(bs3CpuInstr2_sarx_RAX_FSxBX_RCX_icebp), true, 6, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2808	{ BS3_CMN_NM(bs3CpuInstr2_sarx_EAX_EBX_ECX_icebp), false, 5, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2809	{ BS3_CMN_NM(bs3CpuInstr2_sarx_EAX_FSxBX_ECX_icebp), true, 6, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2810	};
2811	return bs3CpuInstr2_Common_Gy_Ey_By(bMode, s_aTests, RT_ELEMENTS(s_aTests), X86_CPUID_STEXT_FEATURE_EBX_BMI1,
2812	0, 0);
2813	}
2814
2815
2816	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_shrx)(uint8_t bMode)
2817	{
2818	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
2819	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests64[] =
2820	{
2821	{ 0, 0, /* -> */ 0, 0 },
2822	{ 0, ~(RTCCUINTXREG)3, /* -> */ 0, 0 },
2823	{ (RTCCUINTXREG)1 << (RTCCINTXREG_BITS - 1), RTCCINTXREG_BITS - 1, /* -> */ 1, 0 },
2824	{ (RTCCUINTXREG)1 << (RTCCINTXREG_BITS - 1), RTCCINTXREG_BITS - 1 + 64, /* -> */ 1, 0 },
2825	{ (RTCCUINTXREG)1 << (RTCCINTXREG_BITS - 2), RTCCINTXREG_BITS - 3, /* -> */ 2, 0 },
2826	{ (RTCCUINTXREG)1 << (RTCCINTXREG_BITS - 2), RTCCINTXREG_BITS - 3 + 64, /* -> */ 2, 0 },
2827	};
2828
2829	/* 32-bit register width */
2830	static BS3CPUINSTR2_SUBTEST_Gy_Ey_By_T const s_aSubTests32[] =
2831	{
2832	{ 0, 0, /* -> */ 0, 0 },
2833	{ 0, ~(RTCCUINTXREG)9, /* -> */ 0, 0 },
2834	{ ~(RTCCUINTXREG)UINT32_C(0x7fffffff), 24, /* -> */ UINT32_C(0x80), 0 },
2835	{ ~(RTCCUINTXREG)UINT32_C(0x7fffffff), 24+32, /* -> */ UINT32_C(0x80), 0 },
2836	{ ~(RTCCUINTXREG)UINT32_C(0xbfffffff), 24, /* -> */ UINT32_C(0x40), 0 },
2837	{ ~(RTCCUINTXREG)UINT32_C(0xbfffffff), 24+32, /* -> */ UINT32_C(0x40), 0 },
2838	};
2839
2840	static BS3CPUINSTR2_TEST_Gy_Ey_By_T const s_aTests[] =
2841	{
2842	{ BS3_CMN_NM(bs3CpuInstr2_shrx_RAX_RBX_RCX_icebp), false, 5, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2843	{ BS3_CMN_NM(bs3CpuInstr2_shrx_RAX_FSxBX_RCX_icebp), true, 6, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
2844	{ BS3_CMN_NM(bs3CpuInstr2_shrx_EAX_EBX_ECX_icebp), false, 5, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2845	{ BS3_CMN_NM(bs3CpuInstr2_shrx_EAX_FSxBX_ECX_icebp), true, 6, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
2846	};
2847	return bs3CpuInstr2_Common_Gy_Ey_By(bMode, s_aTests, RT_ELEMENTS(s_aTests), X86_CPUID_STEXT_FEATURE_EBX_BMI1,
2848	0, 0);
2849	}
2850
2851
2852	/*
2853	* For testing BLSR, BLSMSK, and BLSI.
2854	*/
2855	typedef struct BS3CPUINSTR2_SUBTEST_By_Ey_T
2856	{
2857	RTCCUINTXREG uSrc;
2858	RTCCUINTXREG uDst;
2859	uint16_t fEflOut;
2860	} BS3CPUINSTR2_SUBTEST_By_Ey_T;
2861
2862	typedef struct BS3CPUINSTR2_TEST_By_Ey_T
2863	{
2864	FPFNBS3FAR pfnWorker;
2865	bool fMemSrc;
2866	uint8_t cbInstr;
2867	uint8_t cSubTests;
2868	BS3CPUINSTR2_SUBTEST_By_Ey_T const *paSubTests;
2869	} BS3CPUINSTR2_TEST_By_Ey_T;
2870
2871	static uint8_t bs3CpuInstr2_Common_By_Ey(uint8_t bMode, BS3CPUINSTR2_TEST_By_Ey_T const *paTests, unsigned cTests,
2872	uint32_t fStdExtFeatEbx, uint16_t fEflCheck, uint16_t fEflIgnore)
2873	{
2874	BS3REGCTX Ctx;
2875	BS3TRAPFRAME TrapFrame;
2876	unsigned i, j, k;
2877	uint32_t uStdExtFeatEbx = 0;
2878	bool fSupportsInstr;
2879
2880	fEflCheck &= ~fEflIgnore;
2881
2882	if (g_uBs3CpuDetected & BS3CPU_F_CPUID)
2883	ASMCpuIdExSlow(7, 0, 0, 0, NULL, &uStdExtFeatEbx, NULL, NULL);
2884	fSupportsInstr = RT_BOOL(uStdExtFeatEbx & fStdExtFeatEbx);
2885
2886	/* Ensure the structures are allocated before we sample the stack pointer. */
2887	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
2888	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
2889
2890	/*
2891	* Create test context.
2892	*/
2893	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
2894
2895	/*
2896	* Do the tests twice, first with all flags set, then once again with
2897	* flags cleared. The flags are not supposed to be touched at all.
2898	*/
2899	Ctx.rflags.u16 \|= X86_EFL_STATUS_BITS;
2900	for (j = 0; j < 2; j++)
2901	{
2902	for (i = 0; i < cTests; i++)
2903	{
2904	for (k = 0; k < paTests[i].cSubTests; k++)
2905	{
2906	bool const fOkay = !BS3_MODE_IS_RM_OR_V86(bMode) && fSupportsInstr;
2907	uint8_t const bExpectXcpt = fOkay ? X86_XCPT_DB : X86_XCPT_UD;
2908	uint64_t uExpectRax, uExpectRip;
2909	RTCCUINTXREG uMemSrc, uMemSrcExpect;
2910
2911	Ctx.rax.uCcXReg = ~paTests[i].paSubTests[k].uSrc ^ 0x593e7591;
2912	if (!paTests[i].fMemSrc)
2913	{
2914	Ctx.rbx.uCcXReg = paTests[i].paSubTests[k].uSrc;
2915	uMemSrcExpect = uMemSrc = ~paTests[i].paSubTests[k].uSrc;
2916	}
2917	else
2918	{
2919	uMemSrcExpect = uMemSrc = paTests[i].paSubTests[k].uSrc;
2920	Bs3RegCtxSetGrpSegFromCurPtr(&Ctx, &Ctx.rbx, &Ctx.fs, &uMemSrc);
2921	}
2922	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, paTests[i].pfnWorker);
2923	uExpectRax = fOkay ? paTests[i].paSubTests[k].uDst : Ctx.rax.u;
2924	uExpectRip = Ctx.rip.u + (fOkay ? paTests[i].cbInstr + 1 : 0);
2925	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
2926
2927	if ( TrapFrame.bXcpt != bExpectXcpt
2928	\|\| TrapFrame.Ctx.rip.u != uExpectRip
2929	\|\| TrapFrame.Ctx.rbx.u != Ctx.rbx.u
2930	\|\| TrapFrame.Ctx.rax.u != uExpectRax
2931	/* check that nothing else really changed: */
2932	\|\| (TrapFrame.Ctx.rflags.u16 & fEflCheck)
2933	!= ((fOkay ? paTests[i].paSubTests[k].fEflOut : Ctx.rflags.u16) & fEflCheck)
2934	\|\| (TrapFrame.Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS)
2935	!= (Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS)
2936	\|\| TrapFrame.Ctx.rcx.u != Ctx.rcx.u
2937	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
2938	\|\| TrapFrame.Ctx.rsp.u != Ctx.rsp.u
2939	\|\| TrapFrame.Ctx.rbp.u != Ctx.rbp.u
2940	\|\| TrapFrame.Ctx.rsi.u != Ctx.rsi.u
2941	\|\| TrapFrame.Ctx.rdi.u != Ctx.rdi.u
2942	\|\| uMemSrc != uMemSrcExpect
2943	)
2944	{
2945	Bs3TestFailedF("test #%i/%i failed: input %#" RTCCUINTXREG_XFMT,
2946	i, k, paTests[i].paSubTests[k].uSrc);
2947	if (TrapFrame.bXcpt != bExpectXcpt)
2948	Bs3TestFailedF("Expected bXcpt = %#x, got %#x", bExpectXcpt, TrapFrame.bXcpt);
2949	if (TrapFrame.Ctx.rip.u != uExpectRip)
2950	Bs3TestFailedF("Expected RIP = %#06RX64, got %#06RX64", uExpectRip, TrapFrame.Ctx.rip.u);
2951	if (TrapFrame.Ctx.rax.u != uExpectRax)
2952	Bs3TestFailedF("Expected RAX = %#06RX64, got %#06RX64", uExpectRax, TrapFrame.Ctx.rax.u);
2953	if (TrapFrame.Ctx.rbx.u != Ctx.rbx.u)
2954	Bs3TestFailedF("Expected RBX = %#06RX64, got %#06RX64 (dst)", Ctx.rbx.u, TrapFrame.Ctx.rbx.u);
2955	if ( (TrapFrame.Ctx.rflags.u16 & fEflCheck)
2956	!= ((fOkay ? paTests[i].paSubTests[k].fEflOut : Ctx.rflags.u16) & fEflCheck))
2957	Bs3TestFailedF("Expected EFLAGS = %#06RX32, got %#06RX32 (output)",
2958	(fOkay ? paTests[i].paSubTests[k].fEflOut : Ctx.rflags.u16) & fEflCheck,
2959	TrapFrame.Ctx.rflags.u16 & fEflCheck);
2960	if ( (TrapFrame.Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS)
2961	!= (Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS))
2962	Bs3TestFailedF("Expected EFLAGS = %#06RX32, got %#06RX32 (immutable)",
2963	Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS,
2964	TrapFrame.Ctx.rflags.u16 & ~(fEflCheck \| fEflIgnore) & X86_EFL_STATUS_BITS);
2965
2966	if (TrapFrame.Ctx.rcx.u != Ctx.rcx.u)
2967	Bs3TestFailedF("Expected RCX = %#06RX64, got %#06RX64", Ctx.rcx.u, TrapFrame.Ctx.rcx.u);
2968	if (TrapFrame.Ctx.rdx.u != Ctx.rdx.u)
2969	Bs3TestFailedF("Expected RDX = %#06RX64, got %#06RX64", Ctx.rdx.u, TrapFrame.Ctx.rdx.u);
2970	if (TrapFrame.Ctx.rsp.u != Ctx.rsp.u)
2971	Bs3TestFailedF("Expected RSP = %#06RX64, got %#06RX64", Ctx.rsp.u, TrapFrame.Ctx.rsp.u);
2972	if (TrapFrame.Ctx.rbp.u != Ctx.rbp.u)
2973	Bs3TestFailedF("Expected RBP = %#06RX64, got %#06RX64", Ctx.rbp.u, TrapFrame.Ctx.rbp.u);
2974	if (TrapFrame.Ctx.rsi.u != Ctx.rsi.u)
2975	Bs3TestFailedF("Expected RSI = %#06RX64, got %#06RX64", Ctx.rsi.u, TrapFrame.Ctx.rsi.u);
2976	if (TrapFrame.Ctx.rdi.u != Ctx.rdi.u)
2977	Bs3TestFailedF("Expected RDI = %#06RX64, got %#06RX64", Ctx.rdi.u, TrapFrame.Ctx.rdi.u);
2978	if (uMemSrc != uMemSrcExpect)
2979	Bs3TestFailedF("Expected uMemSrc = %#06RX64, got %#06RX64", (uint64_t)uMemSrcExpect, (uint64_t)uMemSrc);
2980	}
2981	}
2982	}
2983	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
2984	}
2985
2986	return 0;
2987	}
2988
2989
2990	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_blsr)(uint8_t bMode)
2991	{
2992	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
2993	static BS3CPUINSTR2_SUBTEST_By_Ey_T const s_aSubTests64[] =
2994	{
2995	{ 0, /* -> */ 0, X86_EFL_ZF \| X86_EFL_CF },
2996	{ 1, /* -> */ 0, X86_EFL_ZF },
2997	{ 2, /* -> */ 0, X86_EFL_ZF },
2998	{ 3, /* -> */ 2, 0 },
2999	{ 5, /* -> */ 4, 0 },
3000	{ 6, /* -> */ 4, 0 },
3001	{ 7, /* -> */ 6, 0 },
3002	{ 9, /* -> */ 8, 0 },
3003	{ 10, /* -> */ 8, 0 },
3004	{ ~(RTCCUINTXREG)1, /* -> */ ~(RTCCUINTXREG)3, X86_EFL_SF },
3005	{ (RTCCUINTXREG)3 << (RTCCINTXREG_BITS - 2), /* -> */ (RTCCUINTXREG)2 << (RTCCINTXREG_BITS - 2), X86_EFL_SF },
3006	};
3007
3008	/* 32-bit register width */
3009	static BS3CPUINSTR2_SUBTEST_By_Ey_T const s_aSubTests32[] =
3010	{
3011	{ 0, /* -> */ 0, X86_EFL_ZF \| X86_EFL_CF },
3012	{ 1, /* -> */ 0, X86_EFL_ZF },
3013	{ ~(RTCCUINTXREG)1, /* -> */ UINT32_C(0xfffffffc), X86_EFL_SF },
3014	{ ~(RTCCUINTXREG)0 << 30, /* -> */ UINT32_C(0x80000000), X86_EFL_SF },
3015	};
3016
3017	static BS3CPUINSTR2_TEST_By_Ey_T const s_aTests[] =
3018	{
3019	{ BS3_CMN_NM(bs3CpuInstr2_blsr_RAX_RBX_icebp), false, 5, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
3020	{ BS3_CMN_NM(bs3CpuInstr2_blsr_RAX_FSxBX_icebp), true, 6, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
3021	{ BS3_CMN_NM(bs3CpuInstr2_blsr_EAX_EBX_icebp), false, 5, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
3022	{ BS3_CMN_NM(bs3CpuInstr2_blsr_EAX_FSxBX_icebp), true, 6, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
3023	};
3024	return bs3CpuInstr2_Common_By_Ey(bMode, s_aTests, RT_ELEMENTS(s_aTests), X86_CPUID_STEXT_FEATURE_EBX_BMI1,
3025	X86_EFL_STATUS_BITS, 0);
3026	}
3027
3028
3029	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_blsmsk)(uint8_t bMode)
3030	{
3031	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
3032	static BS3CPUINSTR2_SUBTEST_By_Ey_T const s_aSubTests64[] =
3033	{
3034	{ 0, /* -> */ ~(RTCCUINTXREG)0, X86_EFL_CF \| X86_EFL_SF },
3035	{ 1, /* -> */ 1, 0 },
3036	{ ~(RTCCUINTXREG)1, /* -> */ 3, 0 },
3037	{ (RTCCUINTXREG)3 << (RTCCINTXREG_BITS - 2), /* -> */ ~((RTCCUINTXREG)2 << (RTCCINTXREG_BITS - 2)), 0 },
3038	};
3039
3040	/* 32-bit register width */
3041	static BS3CPUINSTR2_SUBTEST_By_Ey_T const s_aSubTests32[] =
3042	{
3043	{ 0, /* -> */ UINT32_MAX, X86_EFL_CF \| X86_EFL_SF },
3044	{ 1, /* -> */ 1, 0 },
3045	{ ~(RTCCUINTXREG)1, /* -> */ 3, 0 },
3046	{ ~(RTCCUINTXREG)0 << 30, /* -> */ UINT32_C(0x7fffffff), 0},
3047	};
3048
3049	static BS3CPUINSTR2_TEST_By_Ey_T const s_aTests[] =
3050	{
3051	{ BS3_CMN_NM(bs3CpuInstr2_blsmsk_RAX_RBX_icebp), false, 5, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
3052	{ BS3_CMN_NM(bs3CpuInstr2_blsmsk_RAX_FSxBX_icebp), true, 6, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
3053	{ BS3_CMN_NM(bs3CpuInstr2_blsmsk_EAX_EBX_icebp), false, 5, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
3054	{ BS3_CMN_NM(bs3CpuInstr2_blsmsk_EAX_FSxBX_icebp), true, 6, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
3055	};
3056	return bs3CpuInstr2_Common_By_Ey(bMode, s_aTests, RT_ELEMENTS(s_aTests), X86_CPUID_STEXT_FEATURE_EBX_BMI1,
3057	X86_EFL_STATUS_BITS, 0);
3058	}
3059
3060
3061	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_blsi)(uint8_t bMode)
3062	{
3063	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
3064	static BS3CPUINSTR2_SUBTEST_By_Ey_T const s_aSubTests64[] =
3065	{
3066	{ 0, /* -> */ 0, X86_EFL_ZF },
3067	{ 1, /* -> */ 1, X86_EFL_CF },
3068	{ ~(RTCCUINTXREG)1, /* -> */ 2, X86_EFL_CF },
3069	{ (RTCCUINTXREG)3 << (RTCCINTXREG_BITS - 2), /* -> */ (RTCCUINTXREG)1 << (RTCCINTXREG_BITS - 2), X86_EFL_CF },
3070	};
3071
3072	/* 32-bit register width */
3073	static BS3CPUINSTR2_SUBTEST_By_Ey_T const s_aSubTests32[] =
3074	{
3075	{ 0, /* -> */ 0, X86_EFL_ZF },
3076	{ 1, /* -> */ 1, X86_EFL_CF },
3077	{ ~(RTCCUINTXREG)1, /* -> */ 2, X86_EFL_CF },
3078	{ ~(RTCCUINTXREG)0 << 30, /* -> */ UINT32_C(0x40000000), X86_EFL_CF },
3079	};
3080
3081	static BS3CPUINSTR2_TEST_By_Ey_T const s_aTests[] =
3082	{
3083	{ BS3_CMN_NM(bs3CpuInstr2_blsi_RAX_RBX_icebp), false, 5, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
3084	{ BS3_CMN_NM(bs3CpuInstr2_blsi_RAX_FSxBX_icebp), true, 6, RT_ELEMENTS(s_aSubTests64), s_aSubTests64 },
3085	{ BS3_CMN_NM(bs3CpuInstr2_blsi_EAX_EBX_icebp), false, 5, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
3086	{ BS3_CMN_NM(bs3CpuInstr2_blsi_EAX_FSxBX_icebp), true, 6, RT_ELEMENTS(s_aSubTests32), s_aSubTests32 },
3087	};
3088	return bs3CpuInstr2_Common_By_Ey(bMode, s_aTests, RT_ELEMENTS(s_aTests), X86_CPUID_STEXT_FEATURE_EBX_BMI1,
3089	X86_EFL_STATUS_BITS, 0);
3090	}
3091
3092
3093	/*
3094	* MULX (BMI2) - destination registers (/r & vvvv) = r/m * rDX
3095	*/
3096	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_mulx)(uint8_t bMode)
3097	{
3098	static const struct
3099	{
3100	FPFNBS3FAR pfnWorker;
3101	bool fMemSrc;
3102	bool fSameDst;
3103	uint8_t cbInstr;
3104	RTCCUINTXREG uSrc1;
3105	RTCCUINTXREG uSrc2;
3106	RTCCUINTXREG uDst1;
3107	RTCCUINTXREG uDst2;
3108	} s_aTests[] =
3109	{
3110	/* 64 bits register width (32 bits in 32- and 16-bit modes): */
3111	{ BS3_CMN_NM(bs3CpuInstr2_mulx_RAX_RCX_RBX_RDX_icebp), false, false, 5, // #0
3112	0, 0, /* -> */ 0, 0 },
3113	{ BS3_CMN_NM(bs3CpuInstr2_mulx_RAX_RCX_RBX_RDX_icebp), false, false, 5, // #1
3114	~(RTCCUINTXREG)0, ~(RTCCUINTXREG)0, /* -> */ ~(RTCCUINTXREG)1, 1 },
3115	{ BS3_CMN_NM(bs3CpuInstr2_mulx_RCX_RCX_RBX_RDX_icebp), false, true, 5, // #2
3116	~(RTCCUINTXREG)0, ~(RTCCUINTXREG)0, /* -> */ ~(RTCCUINTXREG)1, ~(RTCCUINTXREG)1 },
3117	{ BS3_CMN_NM(bs3CpuInstr2_mulx_RAX_RCX_RBX_RDX_icebp), false, false, 5, // #3
3118	2, 2, /* -> */ 0, 4 },
3119	{ BS3_CMN_NM(bs3CpuInstr2_mulx_RAX_RCX_RBX_RDX_icebp), false, false, 5, // #4
3120	~(RTCCUINTXREG)0, 42, /* -> */ 0x29, ~(RTCCUINTXREG)41 },
3121
3122	{ BS3_CMN_NM(bs3CpuInstr2_mulx_RAX_RCX_FSxBX_RDX_icebp), true, false, 6, // #5
3123	0, 0, /* -> */ 0, 0 },
3124	{ BS3_CMN_NM(bs3CpuInstr2_mulx_RAX_RCX_FSxBX_RDX_icebp), true, false, 6, // #6
3125	~(RTCCUINTXREG)0, ~(RTCCUINTXREG)0, /* -> */ ~(RTCCUINTXREG)1, 1 },
3126	{ BS3_CMN_NM(bs3CpuInstr2_mulx_RAX_RCX_FSxBX_RDX_icebp), true, false, 6, // #7
3127	~(RTCCUINTXREG)0, 42, /* -> */ 0x29, ~(RTCCUINTXREG)41 },
3128
3129	/* 32-bit register width */
3130	{ BS3_CMN_NM(bs3CpuInstr2_mulx_EAX_ECX_EBX_EDX_icebp), false, false, 5, // #8
3131	0, 0, /* -> */ 0, 0 },
3132	{ BS3_CMN_NM(bs3CpuInstr2_mulx_EAX_ECX_EBX_EDX_icebp), false, false, 5, // #9
3133	~(RTCCUINTXREG)0, ~(RTCCUINTXREG)0, /* -> */ ~(uint32_t)1, 1 },
3134	{ BS3_CMN_NM(bs3CpuInstr2_mulx_ECX_ECX_EBX_EDX_icebp), false, true, 5, // #10
3135	~(RTCCUINTXREG)0, ~(RTCCUINTXREG)0, /* -> */ ~(uint32_t)1, ~(uint32_t)1 },
3136	{ BS3_CMN_NM(bs3CpuInstr2_mulx_EAX_ECX_EBX_EDX_icebp), false, false, 5, // #11
3137	2, 2, /* -> */ 0, 4 },
3138	{ BS3_CMN_NM(bs3CpuInstr2_mulx_EAX_ECX_EBX_EDX_icebp), false, false, 5, // #12
3139	~(RTCCUINTXREG)0, 42, /* -> */ 0x29, ~(uint32_t)41 },
3140
3141	{ BS3_CMN_NM(bs3CpuInstr2_mulx_EAX_ECX_FSxBX_EDX_icebp), true, false, 6, // #13
3142	0, 0, /* -> */ 0, 0 },
3143	{ BS3_CMN_NM(bs3CpuInstr2_mulx_EAX_ECX_FSxBX_EDX_icebp), true, false, 6, // #14
3144	~(RTCCUINTXREG)0, ~(RTCCUINTXREG)0, /* -> */ ~(uint32_t)1, 1 },
3145	{ BS3_CMN_NM(bs3CpuInstr2_mulx_EAX_ECX_FSxBX_EDX_icebp), true, false, 6, // #15
3146	~(RTCCUINTXREG)0, 42, /* -> */ 0x29, ~(uint32_t)41 },
3147	};
3148
3149	BS3REGCTX Ctx;
3150	BS3TRAPFRAME TrapFrame;
3151	unsigned i, j;
3152	uint32_t uStdExtFeatEbx = 0;
3153	bool fSupportsAndN;
3154
3155	if (g_uBs3CpuDetected & BS3CPU_F_CPUID)
3156	ASMCpuIdExSlow(7, 0, 0, 0, NULL, &uStdExtFeatEbx, NULL, NULL);
3157	fSupportsAndN = RT_BOOL(uStdExtFeatEbx & X86_CPUID_STEXT_FEATURE_EBX_BMI2);
3158
3159	/* Ensure the structures are allocated before we sample the stack pointer. */
3160	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
3161	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
3162
3163	/*
3164	* Create test context.
3165	*/
3166	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
3167
3168	/*
3169	* Do the tests twice, first with all flags set, then once again with
3170	* flags cleared. The flags are not supposed to be touched at all.
3171	*/
3172	Ctx.rflags.u16 \|= X86_EFL_STATUS_BITS;
3173	for (j = 0; j < 2; j++)
3174	{
3175	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
3176	{
3177	bool const fOkay = !BS3_MODE_IS_RM_OR_V86(bMode) && fSupportsAndN;
3178	uint8_t const bExpectXcpt = fOkay ? X86_XCPT_DB : X86_XCPT_UD;
3179	uint64_t uExpectRax, uExpectRcx, uExpectRip;
3180	RTCCUINTXREG uMemSrc1, uMemSrc1Expect;
3181
3182	Ctx.rax.uCcXReg = RTCCUINTXREG_MAX * 1019;
3183	Ctx.rcx.uCcXReg = RTCCUINTXREG_MAX * 4095;
3184	Ctx.rdx.uCcXReg = s_aTests[i].uSrc2;
3185	if (!s_aTests[i].fMemSrc)
3186	{
3187	Ctx.rbx.uCcXReg = s_aTests[i].uSrc1;
3188	uMemSrc1Expect = uMemSrc1 = ~s_aTests[i].uSrc1;
3189	}
3190	else
3191	{
3192	uMemSrc1Expect = uMemSrc1 = s_aTests[i].uSrc1;
3193	Bs3RegCtxSetGrpSegFromCurPtr(&Ctx, &Ctx.rbx, &Ctx.fs, &uMemSrc1);
3194	}
3195	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, s_aTests[i].pfnWorker);
3196	uExpectRax = fOkay && !s_aTests[i].fSameDst ? s_aTests[i].uDst1 : Ctx.rax.u;
3197	uExpectRcx = fOkay ? s_aTests[i].uDst2 : Ctx.rcx.u;
3198	uExpectRip = Ctx.rip.u + (fOkay ? s_aTests[i].cbInstr + 1 : 0);
3199	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
3200
3201	if ( TrapFrame.bXcpt != bExpectXcpt
3202	\|\| TrapFrame.Ctx.rip.u != uExpectRip
3203	\|\| TrapFrame.Ctx.rbx.u != Ctx.rbx.u
3204	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
3205	\|\| TrapFrame.Ctx.rax.u != uExpectRax
3206	\|\| TrapFrame.Ctx.rcx.u != uExpectRcx
3207	/* check that nothing else really changed: */
3208	\|\| (TrapFrame.Ctx.rflags.u16 & X86_EFL_STATUS_BITS) != (Ctx.rflags.u16 & X86_EFL_STATUS_BITS)
3209	\|\| TrapFrame.Ctx.rsp.u != Ctx.rsp.u
3210	\|\| TrapFrame.Ctx.rbp.u != Ctx.rbp.u
3211	\|\| TrapFrame.Ctx.rsi.u != Ctx.rsi.u
3212	\|\| TrapFrame.Ctx.rdi.u != Ctx.rdi.u
3213	\|\| uMemSrc1 != uMemSrc1Expect
3214	)
3215	{
3216	Bs3TestFailedF("test #%i failed: input %#" RTCCUINTXREG_XFMT ", %#" RTCCUINTXREG_XFMT, i, s_aTests[i].uSrc1, s_aTests[i].uSrc2);
3217	if (TrapFrame.bXcpt != bExpectXcpt)
3218	Bs3TestFailedF("Expected bXcpt = %#x, got %#x", bExpectXcpt, TrapFrame.bXcpt);
3219	if (TrapFrame.Ctx.rip.u != uExpectRip)
3220	Bs3TestFailedF("Expected RIP = %#06RX64, got %#06RX64", uExpectRip, TrapFrame.Ctx.rip.u);
3221	if (TrapFrame.Ctx.rax.u != uExpectRax)
3222	Bs3TestFailedF("Expected RAX = %#06RX64, got %#06RX64", uExpectRax, TrapFrame.Ctx.rax.u);
3223	if (TrapFrame.Ctx.rcx.u != uExpectRcx)
3224	Bs3TestFailedF("Expected RCX = %#06RX64, got %#06RX64", uExpectRcx, TrapFrame.Ctx.rcx.u);
3225	if (TrapFrame.Ctx.rbx.u != Ctx.rbx.u)
3226	Bs3TestFailedF("Expected RBX = %#06RX64, got %#06RX64 (dst)", Ctx.rbx.u, TrapFrame.Ctx.rbx.u);
3227	if (TrapFrame.Ctx.rdx.u != Ctx.rdx.u)
3228	Bs3TestFailedF("Expected RDX = %#06RX64, got %#06RX64 (src)", Ctx.rdx.u, TrapFrame.Ctx.rdx.u);
3229
3230	if ( (TrapFrame.Ctx.rflags.u16 & X86_EFL_STATUS_BITS) != (Ctx.rflags.u16 & X86_EFL_STATUS_BITS))
3231	Bs3TestFailedF("Expected EFLAGS = %#06RX32, got %#06RX32 (immutable)",
3232	Ctx.rflags.u16 & X86_EFL_STATUS_BITS, TrapFrame.Ctx.rflags.u16 & X86_EFL_STATUS_BITS);
3233	if (TrapFrame.Ctx.rsp.u != Ctx.rsp.u)
3234	Bs3TestFailedF("Expected RSP = %#06RX64, got %#06RX64", Ctx.rsp.u, TrapFrame.Ctx.rsp.u);
3235	if (TrapFrame.Ctx.rbp.u != Ctx.rbp.u)
3236	Bs3TestFailedF("Expected RBP = %#06RX64, got %#06RX64", Ctx.rbp.u, TrapFrame.Ctx.rbp.u);
3237	if (TrapFrame.Ctx.rsi.u != Ctx.rsi.u)
3238	Bs3TestFailedF("Expected RSI = %#06RX64, got %#06RX64", Ctx.rsi.u, TrapFrame.Ctx.rsi.u);
3239	if (TrapFrame.Ctx.rdi.u != Ctx.rdi.u)
3240	Bs3TestFailedF("Expected RDI = %#06RX64, got %#06RX64", Ctx.rdi.u, TrapFrame.Ctx.rdi.u);
3241	if (uMemSrc1 != uMemSrc1Expect)
3242	Bs3TestFailedF("Expected uMemSrc1 = %#06RX64, got %#06RX64", (uint64_t)uMemSrc1Expect, (uint64_t)uMemSrc1);
3243	}
3244	}
3245	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
3246	}
3247
3248	return 0;
3249	}
3250
3251
3252	/*
3253	* POPCNT - Intel: POPCNT; AMD: ABM.
3254	*/
3255	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_popcnt)(uint8_t bMode)
3256	{
3257	static const struct
3258	{
3259	FPFNBS3FAR pfnWorker;
3260	bool fMemSrc;
3261	uint8_t cWidth;
3262	uint8_t cbInstr;
3263	RTCCUINTXREG uSrc;
3264	RTCCUINTXREG uDst;
3265	uint16_t fEFlags;
3266	} s_aTests[] =
3267	{
3268	/* 16-bit register width */
3269	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_AX_BX_icebp), false, 16, 4 + (ARCH_BITS != 16), // #0
3270	0, /* -> */ 0, X86_EFL_ZF },
3271	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_AX_BX_icebp), false, 16, 4 + (ARCH_BITS != 16), // #1
3272	~(RTCCUINTXREG)0, /* -> */ 16, 0 },
3273	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_AX_BX_icebp), false, 16, 4 + (ARCH_BITS != 16), // #2
3274	UINT16_C(0xffff), /* -> */ 16, 0 },
3275	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_AX_BX_icebp), false, 16, 4 + (ARCH_BITS != 16), // #3
3276	UINT16_C(0x0304), /* -> */ 3, 0 },
3277	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_AX_FSxBX_icebp), true, 16, 5 + (ARCH_BITS != 16), // #4
3278	UINT16_C(0xd569), /* -> */ 9, 0},
3279	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_AX_FSxBX_icebp), true, 16, 5 + (ARCH_BITS != 16), // #5
3280	0, /* -> */ 0, X86_EFL_ZF },
3281
3282	/* 32-bit register width */
3283	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_EAX_EBX_icebp), false, 32, 4 + (ARCH_BITS == 16), // #6
3284	0, /* -> */ 0, X86_EFL_ZF },
3285	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_EAX_EBX_icebp), false, 32, 4 + (ARCH_BITS == 16), // #7
3286	~(RTCCUINTXREG)0, /* -> */ 32, 0},
3287	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_EAX_EBX_icebp), false, 32, 4 + (ARCH_BITS == 16), // #8
3288	UINT32_C(0x01020304), /* -> */ 5, 0},
3289	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_EAX_FSxBX_icebp), true, 32, 5 + (ARCH_BITS == 16), // #9
3290	0, /* -> */ 0, X86_EFL_ZF },
3291	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_EAX_FSxBX_icebp), true, 32, 5 + (ARCH_BITS == 16), // #10
3292	UINT32_C(0x49760948), /* -> */ 12, 0 },
3293
3294	#if ARCH_BITS == 64
3295	/* 64-bit register width */
3296	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_RAX_RBX_icebp), false, 64, 5, // #11
3297	0, /* -> */ 0, X86_EFL_ZF },
3298	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_RAX_RBX_icebp), false, 64, 5, // #12
3299	~(RTCCUINTXREG)0, /* -> */ 64, 0 },
3300	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_RAX_RBX_icebp), false, 64, 5, // #13
3301	UINT64_C(0x1234123412341234), /* -> / 54, 0 },
3302	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_RAX_FSxBX_icebp), true, 64, 6, // #14
3303	0, /* -> */ 0, X86_EFL_ZF },
3304	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_RAX_FSxBX_icebp), true, 64, 6, // #15
3305	~(RTCCUINTXREG)0, /* -> */ 64, 0 },
3306	{ BS3_CMN_NM(bs3CpuInstr2_popcnt_RAX_FSxBX_icebp), true, 64, 6, // #16
3307	UINT64_C(0x5908760293769087), /* -> */ 26, 0 },
3308	#endif
3309	};
3310
3311	BS3REGCTX Ctx;
3312	BS3TRAPFRAME TrapFrame;
3313	unsigned i, j;
3314	bool const fSupportsPopCnt = (g_uBs3CpuDetected & BS3CPU_F_CPUID)
3315	&& (ASMCpuId_ECX(1) & X86_CPUID_FEATURE_ECX_POPCNT);
3316
3317	/* Ensure the structures are allocated before we sample the stack pointer. */
3318	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
3319	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
3320
3321	/*
3322	* Create test context.
3323	*/
3324	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
3325
3326	/*
3327	* Do the tests twice, first with all flags set, then once again with
3328	* flags cleared. The flags are not supposed to be touched at all.
3329	*/
3330	Ctx.rflags.u16 \|= X86_EFL_STATUS_BITS;
3331	for (j = 0; j < 2; j++)
3332	{
3333	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
3334	{
3335	bool const fOkay = fSupportsPopCnt;
3336	uint8_t const bExpectXcpt = fOkay ? X86_XCPT_DB : X86_XCPT_UD;
3337	uint64_t uExpectRax, uExpectRip;
3338	RTCCUINTXREG uMemSrc, uMemSrcExpect;
3339
3340	Ctx.rax.uCcXReg = RTCCUINTXREG_MAX * 1019;
3341	if (!s_aTests[i].fMemSrc)
3342	{
3343	Ctx.rbx.uCcXReg = s_aTests[i].uSrc;
3344	uMemSrcExpect = uMemSrc = ~s_aTests[i].uSrc;
3345	}
3346	else
3347	{
3348	uMemSrcExpect = uMemSrc = s_aTests[i].uSrc;
3349	Bs3RegCtxSetGrpSegFromCurPtr(&Ctx, &Ctx.rbx, &Ctx.fs, &uMemSrc);
3350	}
3351	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, s_aTests[i].pfnWorker);
3352	uExpectRax = fOkay ? s_aTests[i].uDst : Ctx.rax.u;
3353	if (s_aTests[i].cWidth == 16)
3354	uExpectRax = (uExpectRax & UINT16_MAX) \| (Ctx.rax.u & ~(uint64_t)UINT16_MAX);
3355
3356	uExpectRip = Ctx.rip.u + (fOkay ? s_aTests[i].cbInstr + 1 : 0);
3357	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
3358
3359	if ( TrapFrame.bXcpt != bExpectXcpt
3360	\|\| TrapFrame.Ctx.rip.u != uExpectRip
3361	\|\| TrapFrame.Ctx.rbx.u != Ctx.rbx.u
3362	\|\| TrapFrame.Ctx.rax.u != uExpectRax
3363	\|\| (TrapFrame.Ctx.rflags.u16 & X86_EFL_STATUS_BITS) != (fOkay ? s_aTests[i].fEFlags : Ctx.rflags.u16)
3364	/* check that nothing else really changed: */
3365	\|\| TrapFrame.Ctx.rcx.u != Ctx.rcx.u
3366	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
3367	\|\| TrapFrame.Ctx.rsp.u != Ctx.rsp.u
3368	\|\| TrapFrame.Ctx.rbp.u != Ctx.rbp.u
3369	\|\| TrapFrame.Ctx.rsi.u != Ctx.rsi.u
3370	\|\| TrapFrame.Ctx.rdi.u != Ctx.rdi.u
3371	\|\| uMemSrc != uMemSrcExpect
3372	)
3373	{
3374	Bs3TestFailedF("test #%i failed: input %#" RTCCUINTXREG_XFMT, i, s_aTests[i].uSrc);
3375	if (TrapFrame.bXcpt != bExpectXcpt)
3376	Bs3TestFailedF("Expected bXcpt = %#x, got %#x", bExpectXcpt, TrapFrame.bXcpt);
3377	if (TrapFrame.Ctx.rip.u != uExpectRip)
3378	Bs3TestFailedF("Expected RIP = %#06RX64, got %#06RX64", uExpectRip, TrapFrame.Ctx.rip.u);
3379	if (TrapFrame.Ctx.rax.u != uExpectRax)
3380	Bs3TestFailedF("Expected RAX = %#06RX64, got %#06RX64", uExpectRax, TrapFrame.Ctx.rax.u);
3381	if (TrapFrame.Ctx.rbx.u != Ctx.rbx.u)
3382	Bs3TestFailedF("Expected RBX = %#06RX64, got %#06RX64 (dst)", Ctx.rbx.u, TrapFrame.Ctx.rbx.u);
3383	if ((TrapFrame.Ctx.rflags.u16 & X86_EFL_STATUS_BITS) != (fOkay ? s_aTests[i].fEFlags : Ctx.rflags.u16))
3384	Bs3TestFailedF("Expected EFLAGS = %#06RX32, got %#06RX32",
3385	fOkay ? s_aTests[i].fEFlags : Ctx.rflags.u16, TrapFrame.Ctx.rflags.u16 & X86_EFL_STATUS_BITS);
3386
3387	if (TrapFrame.Ctx.rcx.u != Ctx.rcx.u)
3388	Bs3TestFailedF("Expected RCX = %#06RX64, got %#06RX64", Ctx.rcx.u, TrapFrame.Ctx.rcx.u);
3389	if (TrapFrame.Ctx.rdx.u != Ctx.rdx.u)
3390	Bs3TestFailedF("Expected RDX = %#06RX64, got %#06RX64 (src)", Ctx.rdx.u, TrapFrame.Ctx.rdx.u);
3391	if (TrapFrame.Ctx.rsp.u != Ctx.rsp.u)
3392	Bs3TestFailedF("Expected RSP = %#06RX64, got %#06RX64", Ctx.rsp.u, TrapFrame.Ctx.rsp.u);
3393	if (TrapFrame.Ctx.rbp.u != Ctx.rbp.u)
3394	Bs3TestFailedF("Expected RBP = %#06RX64, got %#06RX64", Ctx.rbp.u, TrapFrame.Ctx.rbp.u);
3395	if (TrapFrame.Ctx.rsi.u != Ctx.rsi.u)
3396	Bs3TestFailedF("Expected RSI = %#06RX64, got %#06RX64", Ctx.rsi.u, TrapFrame.Ctx.rsi.u);
3397	if (TrapFrame.Ctx.rdi.u != Ctx.rdi.u)
3398	Bs3TestFailedF("Expected RDI = %#06RX64, got %#06RX64", Ctx.rdi.u, TrapFrame.Ctx.rdi.u);
3399	if (uMemSrc != uMemSrcExpect)
3400	Bs3TestFailedF("Expected uMemSrc = %#06RX64, got %#06RX64", (uint64_t)uMemSrcExpect, (uint64_t)uMemSrc);
3401	}
3402	}
3403	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
3404	}
3405
3406	return 0;
3407	}
3408
3409	/*
3410	* CRC32 - SSE4.2
3411	*/
3412	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_crc32)(uint8_t bMode)
3413	{
3414	typedef struct BS3CPUINSTR2_CRC32_VALUES_T
3415	{
3416	uint32_t uDstIn;
3417	uint32_t uDstOut;
3418	uint64_t uSrc;
3419	} BS3CPUINSTR2_CRC32_VALUES_T;
3420	static const BS3CPUINSTR2_CRC32_VALUES_T s_aValues1[] =
3421	{
3422	{ UINT32_C(0000000000), UINT32_C(0000000000), UINT8_C(0000) },
3423	{ UINT32_C(0xffffffff), UINT32_C(0x25502c8c), UINT8_C(0xea) },
3424	{ UINT32_C(0x25502c8c), UINT32_C(0x474224a6), UINT8_C(0xea) },
3425	{ UINT32_C(0x474224a6), UINT32_C(0x0c7f9048), UINT8_C(0xea) },
3426	{ UINT32_C(0x0c7f9048), UINT32_C(0x39c5b9e0), UINT8_C(0x01) },
3427	{ UINT32_C(0x39c5b9e0), UINT32_C(0x2493fabc), UINT8_C(0x04) },
3428	{ UINT32_C(0x2493fabc), UINT32_C(0x0b05c4d6), UINT8_C(0x27) },
3429	{ UINT32_C(0x0b05c4d6), UINT32_C(0xbe26a561), UINT8_C(0x2a) },
3430	{ UINT32_C(0xbe26a561), UINT32_C(0xe1855652), UINT8_C(0x63) },
3431	{ UINT32_C(0xe1855652), UINT32_C(0xc67efe3f), UINT8_C(0xa7) },
3432	{ UINT32_C(0xc67efe3f), UINT32_C(0x227028cd), UINT8_C(0xfd) },
3433	{ UINT32_C(0x227028cd), UINT32_C(0xf4559a1d), UINT8_C(0xea) },
3434	};
3435	static const BS3CPUINSTR2_CRC32_VALUES_T s_aValues2[] =
3436	{
3437	{ UINT32_C(0000000000), UINT32_C(0000000000), UINT16_C(000000) },
3438	{ UINT32_C(0xffffffff), UINT32_C(0xd550e2a0), UINT16_C(0x04d2) },
3439	{ UINT32_C(0xd550e2a0), UINT32_C(0x38e07a0a), UINT16_C(0xe8cc) },
3440	{ UINT32_C(0x38e07a0a), UINT32_C(0x60ebd519), UINT16_C(0x82a2) },
3441	{ UINT32_C(0x60ebd519), UINT32_C(0xaaa127b5), UINT16_C(0x0fff) },
3442	{ UINT32_C(0xaaa127b5), UINT32_C(0xb13175c6), UINT16_C(0x00ff) },
3443	{ UINT32_C(0xb13175c6), UINT32_C(0x3a226f1b), UINT16_C(0x0300) },
3444	{ UINT32_C(0x3a226f1b), UINT32_C(0xbaedef0c), UINT16_C(0x270f) },
3445	{ UINT32_C(0xbaedef0c), UINT32_C(0x2d18866e), UINT16_C(0x3ff6) },
3446	{ UINT32_C(0x2d18866e), UINT32_C(0x07e2e954), UINT16_C(0x9316) },
3447	{ UINT32_C(0x07e2e954), UINT32_C(0x95f82acb), UINT16_C(0xa59c) },
3448	};
3449	static const BS3CPUINSTR2_CRC32_VALUES_T s_aValues4[] =
3450	{
3451	{ UINT32_C(0000000000), UINT32_C(0000000000), UINT32_C(0000000000) },
3452	{ UINT32_C(0xffffffff), UINT32_C(0xc9a7250e), UINT32_C(0x0270fa68) },
3453	{ UINT32_C(0xc9a7250e), UINT32_C(0x7340d175), UINT32_C(0x23729736) },
3454	{ UINT32_C(0x7340d175), UINT32_C(0x7e17b67d), UINT32_C(0x8bc75d35) },
3455	{ UINT32_C(0x7e17b67d), UINT32_C(0x5028eb71), UINT32_C(0x0e9bebf2) },
3456	{ UINT32_C(0x5028eb71), UINT32_C(0xc0a7f45a), UINT32_C(0x000001bc) },
3457	{ UINT32_C(0xc0a7f45a), UINT32_C(0xa96f4012), UINT32_C(0x0034ba02) },
3458	{ UINT32_C(0xa96f4012), UINT32_C(0xb27c0718), UINT32_C(0x0000002a) },
3459	{ UINT32_C(0xb27c0718), UINT32_C(0x79fb2d35), UINT32_C(0x0153158e) },
3460	{ UINT32_C(0x79fb2d35), UINT32_C(0x23434fc9), UINT32_C(0x02594882) },
3461	{ UINT32_C(0x23434fc9), UINT32_C(0x354bf3b6), UINT32_C(0xb230b8f3) },
3462	};
3463	#if ARCH_BITS >= 64
3464	static const BS3CPUINSTR2_CRC32_VALUES_T s_aValues8[] =
3465	{
3466	{ UINT32_C(0000000000), UINT32_C(0000000000), UINT64_C(000000000000000000) },
3467	{ UINT32_C(0xffffffff), UINT32_C(0xadc36834), UINT64_C(0x02b0b5e2a975c1cc) },
3468	{ UINT32_C(0xadc36834), UINT32_C(0xf0e893c9), UINT64_C(0x823d386bf7517583) },
3469	{ UINT32_C(0xf0e893c9), UINT32_C(0x1a22a837), UINT64_C(0x0481f5311fa061d0) },
3470	{ UINT32_C(0x1a22a837), UINT32_C(0xcf8b6d61), UINT64_C(0x13fa70f64d52a92d) },
3471	{ UINT32_C(0xcf8b6d61), UINT32_C(0xc7dde203), UINT64_C(0x3ccc8b035903d3e1) },
3472	{ UINT32_C(0xc7dde203), UINT32_C(0xd42b5823), UINT64_C(0x0000011850ec2fac) },
3473	{ UINT32_C(0xd42b5823), UINT32_C(0x8b1ce49e), UINT64_C(0x0000000000001364) },
3474	{ UINT32_C(0x8b1ce49e), UINT32_C(0x1af31710), UINT64_C(0x000000057840205a) },
3475	{ UINT32_C(0x1af31710), UINT32_C(0xdea35e8b), UINT64_C(0x2e5d93688d9a0bfa) },
3476	{ UINT32_C(0xdea35e8b), UINT32_C(0x594c013a), UINT64_C(0x8ac7230489e7ffff) },
3477	{ UINT32_C(0x594c013a), UINT32_C(0x27b061e5), UINT64_C(0x6bf037ae325f1c71) },
3478	{ UINT32_C(0x27b061e5), UINT32_C(0x3120b5f7), UINT64_C(0x0fffffff34503556) },
3479	};
3480	#endif
3481	static const struct
3482	{
3483	FPFNBS3FAR pfnWorker;
3484	bool fMemSrc;
3485	uint8_t cbOp;
3486	uint8_t cValues;
3487	BS3CPUINSTR2_CRC32_VALUES_T const BS3_FAR *paValues;
3488	} s_aTests[] =
3489	{
3490	/* 8-bit register width */
3491	{ BS3_CMN_NM(bs3CpuInstr2_crc32_EAX_BL_icebp), false, 1, RT_ELEMENTS(s_aValues1), s_aValues1 },
3492	{ BS3_CMN_NM(bs3CpuInstr2_crc32_EAX_byte_FSxBX_icebp), true, 1, RT_ELEMENTS(s_aValues1), s_aValues1 },
3493
3494	/* 16-bit register width */
3495	{ BS3_CMN_NM(bs3CpuInstr2_crc32_EAX_BX_icebp), false, 2, RT_ELEMENTS(s_aValues2), s_aValues2 },
3496	{ BS3_CMN_NM(bs3CpuInstr2_crc32_EAX_word_FSxBX_icebp), true, 2, RT_ELEMENTS(s_aValues2), s_aValues2 },
3497
3498	/* 32-bit register width */
3499	{ BS3_CMN_NM(bs3CpuInstr2_crc32_EAX_EBX_icebp), false, 4, RT_ELEMENTS(s_aValues4), s_aValues4 },
3500	{ BS3_CMN_NM(bs3CpuInstr2_crc32_EAX_dword_FSxBX_icebp), true, 4, RT_ELEMENTS(s_aValues4), s_aValues4 },
3501	#if ARCH_BITS >= 64
3502	/* 32-bit register width */
3503	{ BS3_CMN_NM(bs3CpuInstr2_crc32_EAX_RBX_icebp), false, 8, RT_ELEMENTS(s_aValues8), s_aValues8 },
3504	{ BS3_CMN_NM(bs3CpuInstr2_crc32_EAX_qword_FSxBX_icebp), true, 8, RT_ELEMENTS(s_aValues8), s_aValues8 },
3505	#endif
3506	};
3507
3508	BS3REGCTX Ctx;
3509	BS3TRAPFRAME TrapFrame;
3510	unsigned i, j;
3511	bool const fSupportsCrc32 = (g_uBs3CpuDetected & BS3CPU_F_CPUID)
3512	&& (ASMCpuId_ECX(1) & X86_CPUID_FEATURE_ECX_SSE4_2);
3513
3514	/* Ensure the structures are allocated before we sample the stack pointer. */
3515	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
3516	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
3517
3518	/*
3519	* Create test context.
3520	*/
3521	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
3522
3523	/*
3524	* Do the tests twice, first with all flags set, then once again with
3525	* flags cleared. The flags are not supposed to be touched at all.
3526	*/
3527	Ctx.rflags.u16 \|= X86_EFL_STATUS_BITS;
3528	for (j = 0; j < 2; j++)
3529	{
3530	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
3531	{
3532	uint8_t const cbOp = s_aTests[i].cbOp;
3533	unsigned const cValues = s_aTests[i].cValues;
3534	BS3CPUINSTR2_CRC32_VALUES_T const BS3_FAR *paValues = s_aTests[i].paValues;
3535	unsigned iValue;
3536	bool const fOkay = fSupportsCrc32;
3537	uint8_t const bExpectXcpt = fOkay ? X86_XCPT_DB : X86_XCPT_UD;
3538	uint64_t const uSrcGarbage = ( cbOp == 1 ? UINT64_C(0x03948314d0f03400)
3539	: cbOp == 2 ? UINT64_C(0x03948314d0f00000)
3540	: cbOp == 4 ? UINT64_C(0x0394831000000000) : 0)
3541	& (ARCH_BITS >= 64 ? UINT64_MAX : UINT32_MAX);
3542	uint64_t uExpectRip;
3543
3544	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, s_aTests[i].pfnWorker);
3545	uExpectRip = Ctx.rip.u + (fOkay ? ((uint8_t const BS3_FAR *)s_aTests[i].pfnWorker)[-1] + 1 : 0);
3546
3547	for (iValue = 0; iValue < cValues; iValue++)
3548	{
3549	uint64_t const uExpectRax = fOkay ? paValues[iValue].uDstOut : paValues[iValue].uDstIn;
3550	uint64_t uMemSrc, uMemSrcExpect;
3551
3552	Ctx.rax.uCcXReg = paValues[iValue].uDstIn;
3553	if (!s_aTests[i].fMemSrc)
3554	{
3555	Ctx.rbx.u64 = paValues[iValue].uSrc \| uSrcGarbage;
3556	uMemSrcExpect = uMemSrc = ~(paValues[iValue].uSrc \| uSrcGarbage);
3557	}
3558	else
3559	{
3560	uMemSrcExpect = uMemSrc = paValues[iValue].uSrc \| uSrcGarbage;
3561	Bs3RegCtxSetGrpSegFromCurPtr(&Ctx, &Ctx.rbx, &Ctx.fs, &uMemSrc);
3562	}
3563
3564	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
3565
3566	if ( TrapFrame.bXcpt != bExpectXcpt
3567	\|\| TrapFrame.Ctx.rip.u != uExpectRip
3568	\|\| TrapFrame.Ctx.rbx.u != Ctx.rbx.u
3569	\|\| TrapFrame.Ctx.rax.u != uExpectRax
3570	/* check that nothing else really changed: */
3571	\|\| TrapFrame.Ctx.rflags.u16 != Ctx.rflags.u16
3572	\|\| TrapFrame.Ctx.rcx.u != Ctx.rcx.u
3573	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
3574	\|\| TrapFrame.Ctx.rsp.u != Ctx.rsp.u
3575	\|\| TrapFrame.Ctx.rbp.u != Ctx.rbp.u
3576	\|\| TrapFrame.Ctx.rsi.u != Ctx.rsi.u
3577	\|\| TrapFrame.Ctx.rdi.u != Ctx.rdi.u
3578	\|\| uMemSrc != uMemSrcExpect
3579	)
3580	{
3581	Bs3TestFailedF("test #%i value #%i failed: input %#RX32, %#RX64",
3582	i, iValue, paValues[iValue].uDstIn, paValues[iValue].uSrc);
3583	if (TrapFrame.bXcpt != bExpectXcpt)
3584	Bs3TestFailedF("Expected bXcpt = %#x, got %#x", bExpectXcpt, TrapFrame.bXcpt);
3585	if (TrapFrame.Ctx.rip.u != uExpectRip)
3586	Bs3TestFailedF("Expected RIP = %#06RX64, got %#06RX64", uExpectRip, TrapFrame.Ctx.rip.u);
3587	if (TrapFrame.Ctx.rax.u != uExpectRax)
3588	Bs3TestFailedF("Expected RAX = %#010RX64, got %#010RX64", uExpectRax, TrapFrame.Ctx.rax.u);
3589	if (TrapFrame.Ctx.rbx.u != Ctx.rbx.u)
3590	Bs3TestFailedF("Expected RBX = %#06RX64, got %#06RX64 (dst)", Ctx.rbx.u, TrapFrame.Ctx.rbx.u);
3591
3592	if (TrapFrame.Ctx.rflags.u16 != Ctx.rflags.u16)
3593	Bs3TestFailedF("Expected EFLAGS = %#06RX32, got %#06RX32", Ctx.rflags.u16, TrapFrame.Ctx.rflags.u16);
3594	if (TrapFrame.Ctx.rcx.u != Ctx.rcx.u)
3595	Bs3TestFailedF("Expected RCX = %#06RX64, got %#06RX64", Ctx.rcx.u, TrapFrame.Ctx.rcx.u);
3596	if (TrapFrame.Ctx.rdx.u != Ctx.rdx.u)
3597	Bs3TestFailedF("Expected RDX = %#06RX64, got %#06RX64 (src)", Ctx.rdx.u, TrapFrame.Ctx.rdx.u);
3598	if (TrapFrame.Ctx.rsp.u != Ctx.rsp.u)
3599	Bs3TestFailedF("Expected RSP = %#06RX64, got %#06RX64", Ctx.rsp.u, TrapFrame.Ctx.rsp.u);
3600	if (TrapFrame.Ctx.rbp.u != Ctx.rbp.u)
3601	Bs3TestFailedF("Expected RBP = %#06RX64, got %#06RX64", Ctx.rbp.u, TrapFrame.Ctx.rbp.u);
3602	if (TrapFrame.Ctx.rsi.u != Ctx.rsi.u)
3603	Bs3TestFailedF("Expected RSI = %#06RX64, got %#06RX64", Ctx.rsi.u, TrapFrame.Ctx.rsi.u);
3604	if (TrapFrame.Ctx.rdi.u != Ctx.rdi.u)
3605	Bs3TestFailedF("Expected RDI = %#06RX64, got %#06RX64", Ctx.rdi.u, TrapFrame.Ctx.rdi.u);
3606	if (uMemSrc != uMemSrcExpect)
3607	Bs3TestFailedF("Expected uMemSrc = %#06RX64, got %#06RX64", (uint64_t)uMemSrcExpect, (uint64_t)uMemSrc);
3608	}
3609	}
3610	}
3611	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
3612	}
3613
3614	return 0;
3615	}
3616
3617	#if 0 /* Program for generating CRC32 value sets: */
3618	#include <stdio.h>
3619	#include <stdint.h>
3620	#include <stdlib.h>
3621
3622	int main(int argc, char **argv)
3623	{
3624	int cbOp = atoi(argv[1]);
3625	uint32_t uBefore = atoi(argv[2]);
3626	int i = 3;
3627	while (i < argc)
3628	{
3629	unsigned long long uValue = strtoull(argv[i], NULL, 0);
3630	uint32_t uAfter = uBefore;
3631	switch (cbOp)
3632	{
3633	case 1:
3634	__asm__ __volatile__("crc32b %2, %0" : "=r" (uAfter) : "0" (uAfter), "r" ((uint8_t)uValue));
3635	printf(" { UINT32_C(%#010x), UINT32_C(%#010x), UINT8_C(%#04x) },\n",
3636	uBefore, uAfter, (unsigned)(uint8_t)uValue);
3637	break;
3638	case 2:
3639	__asm__ __volatile__("crc32w %2, %0" : "=r" (uAfter) : "0" (uAfter), "r" ((uint16_t)uValue));
3640	printf(" { UINT32_C(%#010x), UINT32_C(%#010x), UINT16_C(%#06x) },\n",
3641	uBefore, uAfter, (unsigned)(uint16_t)uValue);
3642	break;
3643	case 4:
3644	__asm__ __volatile__("crc32l %2, %0" : "=r" (uAfter) : "0" (uAfter), "r" ((uint32_t)uValue));
3645	printf(" { UINT32_C(%#010x), UINT32_C(%#010x), UINT32_C(%#010x) },\n",
3646	uBefore, uAfter, (uint32_t)uValue);
3647	break;
3648	case 8:
3649	{
3650	uint64_t u64After = uBefore;
3651	__asm__ __volatile__("crc32q %2, %0" : "=r" (u64After) : "0" (u64After), "r" (uValue));
3652	uAfter = (uint32_t)u64After;
3653	printf(" { UINT32_C(%#010x), UINT32_C(%#010x), UINT64_C(%#018llx) },\n", uBefore, uAfter, uValue);
3654	break;
3655	}
3656	}
3657
3658	/* next */
3659	uBefore = uAfter;
3660	i++;
3661	}
3662	return 0;
3663	}
3664	#endif
3665
3666
3667	/*
3668	* ADCX/ADOX - ADX
3669	*/
3670	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_adcx_adox)(uint8_t bMode)
3671	{
3672	typedef struct BS3CPUINSTR2_ADX_VALUES_T
3673	{
3674	uint64_t uDstOut;
3675	uint64_t uDstIn;
3676	uint64_t uSrc;
3677	bool fFlagIn;
3678	bool fFlagOut;
3679	} BS3CPUINSTR2_ADX_VALUES_T;
3680	static const BS3CPUINSTR2_ADX_VALUES_T s_aValues4[] =
3681	{
3682	{ UINT32_C(0000000000), UINT32_C(0000000000), UINT32_C(0000000000), false, false },
3683	{ UINT32_C(0000000001), UINT32_C(0000000000), UINT32_C(0000000000), true, false },
3684
3685	{ UINT32_C(0xfffffffe), UINT32_MAX / 2, UINT32_MAX / 2, false, false },
3686	{ UINT32_C(0xffffffff), UINT32_MAX / 2, UINT32_MAX / 2, true, false },
3687
3688	{ UINT32_C(0x7ffffffe), UINT32_MAX, UINT32_MAX / 2, false, true },
3689	{ UINT32_C(0x7fffffff), UINT32_MAX, UINT32_MAX / 2, true, true },
3690
3691	{ UINT32_C(0x7ffffffe), UINT32_MAX / 2, UINT32_MAX, false, true },
3692	{ UINT32_C(0x7fffffff), UINT32_MAX / 2, UINT32_MAX, true, true },
3693
3694	{ UINT32_C(0xfffffffe), UINT32_MAX, UINT32_MAX, false, true },
3695	{ UINT32_C(0xffffffff), UINT32_MAX, UINT32_MAX, true, true },
3696	};
3697	#if ARCH_BITS >= 64
3698	static const BS3CPUINSTR2_ADX_VALUES_T s_aValues8[] =
3699	{
3700	{ UINT64_C(00000000000000000000), UINT64_C(00000000000000000000), UINT64_C(00000000000000000000), false, false },
3701	{ UINT64_C(00000000000000000001), UINT64_C(00000000000000000000), UINT64_C(00000000000000000000), true, false },
3702
3703	{ UINT64_C(0xfffffffffffffffe), UINT64_MAX / 2, UINT64_MAX / 2, false, false },
3704	{ UINT64_C(0xffffffffffffffff), UINT64_MAX / 2, UINT64_MAX / 2, true, false },
3705
3706	{ UINT64_C(0x7ffffffffffffffe), UINT64_MAX, UINT64_MAX / 2, false, true },
3707	{ UINT64_C(0x7fffffffffffffff), UINT64_MAX, UINT64_MAX / 2, true, true },
3708
3709	{ UINT64_C(0x7ffffffffffffffe), UINT64_MAX / 2, UINT64_MAX, false, true },
3710	{ UINT64_C(0x7fffffffffffffff), UINT64_MAX / 2, UINT64_MAX, true, true },
3711
3712	{ UINT64_C(0xfffffffffffffffe), UINT64_MAX, UINT64_MAX, false, true },
3713	{ UINT64_C(0xffffffffffffffff), UINT64_MAX, UINT64_MAX, true, true },
3714	};
3715	#endif
3716	static const struct
3717	{
3718	FPFNBS3FAR pfnWorker;
3719	bool fMemSrc;
3720	uint8_t cbOp;
3721	uint8_t cValues;
3722	BS3CPUINSTR2_ADX_VALUES_T const BS3_FAR *paValues;
3723	uint32_t fEFlagsMod;
3724	} s_aTests[] =
3725	{
3726	/* 32-bit register width */
3727	{ BS3_CMN_NM(bs3CpuInstr2_adcx_EAX_EBX_icebp), false, 4, RT_ELEMENTS(s_aValues4), s_aValues4, X86_EFL_CF },
3728	{ BS3_CMN_NM(bs3CpuInstr2_adcx_EAX_dword_FSxBX_icebp), true, 4, RT_ELEMENTS(s_aValues4), s_aValues4, X86_EFL_CF },
3729
3730	{ BS3_CMN_NM(bs3CpuInstr2_adox_EAX_EBX_icebp), false, 4, RT_ELEMENTS(s_aValues4), s_aValues4, X86_EFL_OF },
3731	{ BS3_CMN_NM(bs3CpuInstr2_adox_EAX_dword_FSxBX_icebp), true, 4, RT_ELEMENTS(s_aValues4), s_aValues4, X86_EFL_OF },
3732	#if ARCH_BITS >= 64
3733	/* 64-bit register width */
3734	{ BS3_CMN_NM(bs3CpuInstr2_adcx_RAX_RBX_icebp), false, 8, RT_ELEMENTS(s_aValues8), s_aValues8, X86_EFL_CF },
3735	{ BS3_CMN_NM(bs3CpuInstr2_adcx_RAX_qword_FSxBX_icebp), true, 8, RT_ELEMENTS(s_aValues8), s_aValues8, X86_EFL_CF },
3736
3737	{ BS3_CMN_NM(bs3CpuInstr2_adox_RAX_RBX_icebp), false, 8, RT_ELEMENTS(s_aValues8), s_aValues8, X86_EFL_OF },
3738	{ BS3_CMN_NM(bs3CpuInstr2_adox_RAX_qword_FSxBX_icebp), true, 8, RT_ELEMENTS(s_aValues8), s_aValues8, X86_EFL_OF },
3739	#endif
3740	};
3741
3742	BS3REGCTX Ctx;
3743	BS3TRAPFRAME TrapFrame;
3744	unsigned i, j;
3745	bool fSupportsAdx = false;
3746
3747	if ( (g_uBs3CpuDetected & BS3CPU_F_CPUID)
3748	&& ASMCpuId_EAX(0) >= 7)
3749	{
3750	uint32_t fEbx = 0;
3751	ASMCpuIdExSlow(7, 0, 0, 0, NULL, &fEbx, NULL, NULL);
3752	fSupportsAdx = RT_BOOL(fEbx & X86_CPUID_STEXT_FEATURE_EBX_ADX);
3753	}
3754
3755	/* Ensure the structures are allocated before we sample the stack pointer. */
3756	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
3757	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
3758
3759	/*
3760	* Create test context.
3761	*/
3762	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
3763
3764	/*
3765	* Do the tests twice, first with all flags set, then once again with
3766	* flags cleared. The flags are not supposed to be touched at all except for the one indicated (CF or OF).
3767	*/
3768	Ctx.rflags.u16 \|= X86_EFL_STATUS_BITS;
3769	for (j = 0; j < 2; j++)
3770	{
3771	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
3772	{
3773	uint8_t const cbOp = s_aTests[i].cbOp;
3774	unsigned const cValues = s_aTests[i].cValues;
3775	BS3CPUINSTR2_ADX_VALUES_T const BS3_FAR *paValues = s_aTests[i].paValues;
3776	uint32_t const fEFlagsMod = s_aTests[i].fEFlagsMod;
3777	unsigned iValue;
3778	bool const fOkay = fSupportsAdx;
3779	uint8_t const bExpectXcpt = fOkay ? X86_XCPT_DB : X86_XCPT_UD;
3780	uint64_t const uSrcGarbage = ( cbOp == 4 ? UINT64_C(0x0394831000000000) : 0)
3781	& (ARCH_BITS >= 64 ? UINT64_MAX : UINT32_MAX);
3782	uint64_t uExpectRip;
3783
3784	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, s_aTests[i].pfnWorker);
3785	uExpectRip = Ctx.rip.u + (fOkay ? ((uint8_t const BS3_FAR *)s_aTests[i].pfnWorker)[-1] + 1 : 0);
3786
3787	for (iValue = 0; iValue < cValues; iValue++)
3788	{
3789	uint64_t const uExpectRax = fOkay ? paValues[iValue].uDstOut : paValues[iValue].uDstIn;
3790	uint64_t uMemSrc, uMemSrcExpect;
3791
3792	Ctx.rax.uCcXReg = paValues[iValue].uDstIn;
3793	if (!s_aTests[i].fMemSrc)
3794	{
3795	Ctx.rbx.u64 = paValues[iValue].uSrc \| uSrcGarbage;
3796	uMemSrcExpect = uMemSrc = ~(paValues[iValue].uSrc \| uSrcGarbage);
3797	}
3798	else
3799	{
3800	uMemSrcExpect = uMemSrc = paValues[iValue].uSrc \| uSrcGarbage;
3801	Bs3RegCtxSetGrpSegFromCurPtr(&Ctx, &Ctx.rbx, &Ctx.fs, &uMemSrc);
3802	}
3803
3804	Ctx.rflags.u16 &= ~fEFlagsMod;
3805	if (paValues[iValue].fFlagIn)
3806	Ctx.rflags.u16 \|= fEFlagsMod;
3807
3808	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
3809
3810	if (fOkay)
3811	{
3812	Ctx.rflags.u16 &= ~fEFlagsMod;
3813	if (paValues[iValue].fFlagOut)
3814	Ctx.rflags.u16 \|= fEFlagsMod;
3815	}
3816
3817	if ( TrapFrame.bXcpt != bExpectXcpt
3818	\|\| TrapFrame.Ctx.rip.u != uExpectRip
3819	\|\| TrapFrame.Ctx.rbx.u != Ctx.rbx.u
3820	\|\| TrapFrame.Ctx.rax.u != uExpectRax
3821	/* check that nothing else really changed: */
3822	\|\| TrapFrame.Ctx.rflags.u16 != Ctx.rflags.u16
3823	\|\| TrapFrame.Ctx.rcx.u != Ctx.rcx.u
3824	\|\| TrapFrame.Ctx.rdx.u != Ctx.rdx.u
3825	\|\| TrapFrame.Ctx.rsp.u != Ctx.rsp.u
3826	\|\| TrapFrame.Ctx.rbp.u != Ctx.rbp.u
3827	\|\| TrapFrame.Ctx.rsi.u != Ctx.rsi.u
3828	\|\| TrapFrame.Ctx.rdi.u != Ctx.rdi.u
3829	\|\| uMemSrc != uMemSrcExpect
3830	)
3831	{
3832	Bs3TestFailedF("test #%i value #%i failed: input %#RX64, %#RX64",
3833	i, iValue, paValues[iValue].uDstIn, paValues[iValue].uSrc);
3834	if (TrapFrame.bXcpt != bExpectXcpt)
3835	Bs3TestFailedF("Expected bXcpt = %#x, got %#x", bExpectXcpt, TrapFrame.bXcpt);
3836	if (TrapFrame.Ctx.rip.u != uExpectRip)
3837	Bs3TestFailedF("Expected RIP = %#06RX64, got %#06RX64", uExpectRip, TrapFrame.Ctx.rip.u);
3838	if (TrapFrame.Ctx.rax.u != uExpectRax)
3839	Bs3TestFailedF("Expected RAX = %#010RX64, got %#010RX64", uExpectRax, TrapFrame.Ctx.rax.u);
3840	if (TrapFrame.Ctx.rbx.u != Ctx.rbx.u)
3841	Bs3TestFailedF("Expected RBX = %#06RX64, got %#06RX64 (dst)", Ctx.rbx.u, TrapFrame.Ctx.rbx.u);
3842
3843	if (TrapFrame.Ctx.rflags.u16 != Ctx.rflags.u16)
3844	Bs3TestFailedF("Expected EFLAGS = %#06RX16, got %#06RX16", Ctx.rflags.u16, TrapFrame.Ctx.rflags.u16);
3845	if (TrapFrame.Ctx.rcx.u != Ctx.rcx.u)
3846	Bs3TestFailedF("Expected RCX = %#06RX64, got %#06RX64", Ctx.rcx.u, TrapFrame.Ctx.rcx.u);
3847	if (TrapFrame.Ctx.rdx.u != Ctx.rdx.u)
3848	Bs3TestFailedF("Expected RDX = %#06RX64, got %#06RX64 (src)", Ctx.rdx.u, TrapFrame.Ctx.rdx.u);
3849	if (TrapFrame.Ctx.rsp.u != Ctx.rsp.u)
3850	Bs3TestFailedF("Expected RSP = %#06RX64, got %#06RX64", Ctx.rsp.u, TrapFrame.Ctx.rsp.u);
3851	if (TrapFrame.Ctx.rbp.u != Ctx.rbp.u)
3852	Bs3TestFailedF("Expected RBP = %#06RX64, got %#06RX64", Ctx.rbp.u, TrapFrame.Ctx.rbp.u);
3853	if (TrapFrame.Ctx.rsi.u != Ctx.rsi.u)
3854	Bs3TestFailedF("Expected RSI = %#06RX64, got %#06RX64", Ctx.rsi.u, TrapFrame.Ctx.rsi.u);
3855	if (TrapFrame.Ctx.rdi.u != Ctx.rdi.u)
3856	Bs3TestFailedF("Expected RDI = %#06RX64, got %#06RX64", Ctx.rdi.u, TrapFrame.Ctx.rdi.u);
3857	if (uMemSrc != uMemSrcExpect)
3858	Bs3TestFailedF("Expected uMemSrc = %#06RX64, got %#06RX64", (uint64_t)uMemSrcExpect, (uint64_t)uMemSrc);
3859	}
3860	}
3861	}
3862	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
3863	}
3864
3865	return 0;
3866	}
3867
3868
3869
3870	/*
3871	* MOVBE
3872	*/
3873	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_movbe)(uint8_t bMode)
3874	{
3875	const char BS3_FAR * const pszMode = Bs3GetModeName(bMode);
3876
3877	typedef struct BS3CPUINSTR2_MOVBE_VALUES_T
3878	{
3879	RTCCUINTXREG uDstOut;
3880	RTCCUINTXREG uDstIn;
3881	RTCCUINTXREG uSrc;
3882	} BS3CPUINSTR2_MOVBE_VALUES_T;
3883	static const BS3CPUINSTR2_MOVBE_VALUES_T s_aValues2[] =
3884	{
3885	{ UINT64_C(0xc0dedeaddead3412), UINT64_C(0xc0dedeaddeadc0de), UINT16_C(0x1234) }
3886	};
3887	static const BS3CPUINSTR2_MOVBE_VALUES_T s_aValues4MemSrc[] =
3888	{
3889	{ UINT64_C(0x78563412), UINT64_C(0xc0dedeaddeadc0de), UINT32_C(0x12345678) }
3890	};
3891	static const BS3CPUINSTR2_MOVBE_VALUES_T s_aValues4MemDst[] =
3892	{
3893	{ UINT64_C(0xc0dedead78563412), UINT64_C(0xc0dedeaddeadc0de), UINT32_C(0x12345678) }
3894	};
3895	#if ARCH_BITS >= 64
3896	static const BS3CPUINSTR2_MOVBE_VALUES_T s_aValues8[] =
3897	{
3898	{ UINT64_C(0xf0debc9a78563412), UINT64_C(0xc0dedeaddeadc0de), UINT64_C(0x123456789abcdef0) }
3899	};
3900	#endif
3901	static const struct
3902	{
3903	FPFNBS3FAR pfnWorker;
3904	bool fMemSrc;
3905	uint8_t offIcebp;
3906	uint8_t cValues;
3907	BS3CPUINSTR2_MOVBE_VALUES_T const BS3_FAR *paValues;
3908	} s_aTests[] =
3909	{
3910	/* 16-bit register width */
3911	{ BS3_CMN_NM(bs3CpuInstr2_movbe_AX_word_FSxBX_icebp), true, 6 + (ARCH_BITS != 16), RT_ELEMENTS(s_aValues2), s_aValues2 },
3912	{ BS3_CMN_NM(bs3CpuInstr2_movbe_word_FSxBX_AX_icebp), false, 6 + (ARCH_BITS != 16), RT_ELEMENTS(s_aValues2), s_aValues2 },
3913	/* 32-bit register width */
3914	{ BS3_CMN_NM(bs3CpuInstr2_movbe_EAX_dword_FSxBX_icebp), true, 6 + (ARCH_BITS == 16), RT_ELEMENTS(s_aValues4MemSrc), s_aValues4MemSrc },
3915	{ BS3_CMN_NM(bs3CpuInstr2_movbe_dword_FSxBX_EAX_icebp), false, 6 + (ARCH_BITS == 16), RT_ELEMENTS(s_aValues4MemDst), s_aValues4MemDst },
3916	#if ARCH_BITS >= 64
3917	/* 64-bit register width */
3918	{ BS3_CMN_NM(bs3CpuInstr2_movbe_RAX_qword_FSxBX_icebp), true, 7, RT_ELEMENTS(s_aValues8), s_aValues8 },
3919	{ BS3_CMN_NM(bs3CpuInstr2_movbe_qword_FSxBX_RAX_icebp), false, 7, RT_ELEMENTS(s_aValues8), s_aValues8 },
3920	#endif
3921	};
3922
3923	BS3REGCTX Ctx;
3924	BS3REGCTX ExpectCtx;
3925	BS3TRAPFRAME TrapFrame;
3926	unsigned i, j;
3927	bool fSupportsMovBe = false;
3928
3929	if ( (g_uBs3CpuDetected & BS3CPU_F_CPUID)
3930	&& ASMCpuId_EAX(0) >= 1)
3931	{
3932	uint32_t fEcx = 0;
3933	ASMCpuIdExSlow(1, 0, 0, 0, NULL, NULL, &fEcx, NULL);
3934	fSupportsMovBe = RT_BOOL(fEcx & X86_CPUID_FEATURE_ECX_MOVBE);
3935	}
3936
3937	/* Ensure the structures are allocated before we sample the stack pointer. */
3938	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
3939	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
3940	Bs3MemSet(&ExpectCtx, 0, sizeof(ExpectCtx));
3941
3942	/*
3943	* Create test context.
3944	*/
3945	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
3946
3947	/*
3948	* Do the tests twice, first with all flags set, then once again with
3949	* flags cleared. The flags are not supposed to be touched at all.
3950	*/
3951	g_usBs3TestStep = 0;
3952	for (j = 0; j < 2; j++)
3953	{
3954	for (i = 0; i < RT_ELEMENTS(s_aTests); i++)
3955	{
3956	unsigned const cValues = s_aTests[i].cValues;
3957	BS3CPUINSTR2_MOVBE_VALUES_T const BS3_FAR *paValues = s_aTests[i].paValues;
3958	unsigned iValue;
3959	bool const fOkay = fSupportsMovBe;
3960	uint8_t const bExpectXcpt = fOkay ? X86_XCPT_DB : X86_XCPT_UD;
3961	uint64_t uExpectRip;
3962
3963	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, s_aTests[i].pfnWorker);
3964	uExpectRip = Ctx.rip.u + (fOkay ? ((uint8_t const BS3_FAR *)s_aTests[i].pfnWorker)[-1] + 1 : 0);
3965
3966	for (iValue = 0; iValue < cValues; iValue++)
3967	{
3968	//uint64_t const uExpectRax = fOkay ? paValues[iValue].uDstOut : paValues[iValue].uDstIn;
3969	uint64_t uMem, uMemExpect;
3970
3971	Bs3MemCpy(&ExpectCtx, &Ctx, sizeof(ExpectCtx));
3972
3973	if (!s_aTests[i].fMemSrc)
3974	{
3975	/* Memory is destination */
3976	Ctx.rax.u64 = paValues[iValue].uSrc;
3977	ExpectCtx.rax.u64 = paValues[iValue].uSrc;
3978	uMem = paValues[iValue].uDstIn;
3979	uMemExpect = paValues[iValue].uDstOut;
3980	Bs3RegCtxSetGrpSegFromCurPtr(&Ctx, &Ctx.rbx, &Ctx.fs, &uMem);
3981	Bs3RegCtxSetGrpSegFromCurPtr(&ExpectCtx, &ExpectCtx.rbx, &ExpectCtx.fs, &uMem);
3982	}
3983	else
3984	{
3985	/* Memory is source */
3986	uMemExpect = uMem = paValues[iValue].uSrc;
3987	Ctx.rax.u64 = paValues[iValue].uDstIn;
3988	ExpectCtx.rax.u64 = paValues[iValue].uDstOut;
3989	Bs3RegCtxSetGrpSegFromCurPtr(&Ctx, &Ctx.rbx, &Ctx.fs, &uMem);
3990	Bs3RegCtxSetGrpSegFromCurPtr(&ExpectCtx, &ExpectCtx.rbx, &ExpectCtx.fs, &uMem);
3991	}
3992
3993	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
3994	g_usBs3TestStep++;
3995
3996	if ( !Bs3TestCheckRegCtxEx(&TrapFrame.Ctx, &ExpectCtx, bExpectXcpt == X86_XCPT_DB ? s_aTests[i].offIcebp : 0 /cbPcAdjust/,
3997	0 /cbSpAcjust/, 0 /fExtraEfl/, pszMode, g_usBs3TestStep)
3998	\|\| TrapFrame.bXcpt != bExpectXcpt
3999	\|\| uMem != uMemExpect
4000	)
4001	{
4002	if (TrapFrame.bXcpt != bExpectXcpt)
4003	Bs3TestFailedF("Expected bXcpt=#%x, got %#x (%#x)", bExpectXcpt, TrapFrame.bXcpt, TrapFrame.uErrCd);
4004	if (uMem != uMemExpect)
4005	Bs3TestFailedF("Expected uMem = %#06RX64, got %#06RX64", (uint64_t)uMemExpect, (uint64_t)uMem);
4006	Bs3TestFailedF("^^^ iCfg=%u iWorker=%d iValue=%d\n",
4007	j, i, iValue);
4008	}
4009	}
4010	}
4011	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
4012	}
4013
4014	return 0;
4015	}
4016
4017
4018
4019	/*
4020	* CMPXCHG8B
4021	*/
4022	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_cmpxchg8b)(uint8_t bMode)
4023	{
4024
4025	BS3REGCTX Ctx;
4026	BS3REGCTX ExpectCtx;
4027	BS3TRAPFRAME TrapFrame;
4028	RTUINT64U au64[3];
4029	PRTUINT64U pau64 = RT_ALIGN_PT(&au64[0], sizeof(RTUINT64U), PRTUINT64U);
4030	bool const fSupportCX8 = RT_BOOL(ASMCpuId_EDX(1) & X86_CPUID_FEATURE_EDX_CX8);
4031	const char BS3_FAR * const pszMode = Bs3GetModeName(bMode);
4032	uint8_t bRing = BS3_MODE_IS_V86(bMode) ? 3 : 0;
4033	unsigned iFlags;
4034	unsigned offBuf;
4035	unsigned iMatch;
4036	unsigned iWorker;
4037	static struct
4038	{
4039	bool fLocked;
4040	uint8_t offIcebp;
4041	FNBS3FAR *pfnWorker;
4042	} const s_aWorkers[] =
4043	{
4044	{ false, 5, BS3_CMN_NM(bs3CpuInstr2_cmpxchg8b_FSxDI_icebp) },
4045	#if TMPL_MODE == BS3_MODE_RM \|\| TMPL_MODE == BS3_MODE_PP16
4046	{ false, 5, BS3_CMN_NM(bs3CpuInstr2_o16_cmpxchg8b_FSxDI_icebp) },
4047	#else
4048	{ false, 6, BS3_CMN_NM(bs3CpuInstr2_o16_cmpxchg8b_FSxDI_icebp) },
4049	#endif
4050	{ false, 6, BS3_CMN_NM(bs3CpuInstr2_repz_cmpxchg8b_FSxDI_icebp) },
4051	{ false, 6, BS3_CMN_NM(bs3CpuInstr2_repnz_cmpxchg8b_FSxDI_icebp) },
4052	{ true, 1+5, BS3_CMN_NM(bs3CpuInstr2_lock_cmpxchg8b_FSxDI_icebp) },
4053	{ true, 1+6, BS3_CMN_NM(bs3CpuInstr2_lock_o16_cmpxchg8b_FSxDI_icebp) },
4054	{ true, 1+6, BS3_CMN_NM(bs3CpuInstr2_lock_repz_cmpxchg8b_FSxDI_icebp) },
4055	{ true, 1+6, BS3_CMN_NM(bs3CpuInstr2_lock_repnz_cmpxchg8b_FSxDI_icebp) },
4056	};
4057
4058	/* Ensure the structures are allocated before we sample the stack pointer. */
4059	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
4060	Bs3MemSet(&ExpectCtx, 0, sizeof(ExpectCtx));
4061	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
4062	Bs3MemSet(pau64, 0, sizeof(pau64[0]) * 2);
4063
4064	/*
4065	* Create test context.
4066	*/
4067	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
4068	if (!fSupportCX8)
4069	Bs3TestPrintf("Note! CMPXCHG8B is not supported by the CPU!\n");
4070
4071	/*
4072	* Run the tests in all rings since alignment issues may behave
4073	* differently in ring-3 compared to ring-0.
4074	*/
4075	for (;;)
4076	{
4077	/*
4078	* One loop with alignment checks disabled and one with enabled.
4079	*/
4080	unsigned iCfg;
4081	for (iCfg = 0; iCfg < 2; iCfg++)
4082	{
4083	if (iCfg)
4084	{
4085	Ctx.rflags.u32 \|= X86_EFL_AC;
4086	Ctx.cr0.u32 \|= X86_CR0_AM;
4087	}
4088	else
4089	{
4090	Ctx.rflags.u32 &= ~X86_EFL_AC;
4091	Ctx.cr0.u32 &= ~X86_CR0_AM;
4092	}
4093
4094	/*
4095	* One loop with the normal variant and one with the locked one
4096	*/
4097	g_usBs3TestStep = 0;
4098	for (iWorker = 0; iWorker < RT_ELEMENTS(s_aWorkers); iWorker++)
4099	{
4100	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, s_aWorkers[iWorker].pfnWorker);
4101
4102	/*
4103	* One loop with all status flags set, and one with them clear.
4104	*/
4105	Ctx.rflags.u16 \|= X86_EFL_STATUS_BITS;
4106	for (iFlags = 0; iFlags < 2; iFlags++)
4107	{
4108	Bs3MemCpy(&ExpectCtx, &Ctx, sizeof(ExpectCtx));
4109
4110	for (offBuf = 0; offBuf < sizeof(RTUINT64U); offBuf++)
4111	{
4112	# define CX8_OLD_LO UINT32_C(0xcc9c4bbd)
4113	# define CX8_OLD_HI UINT32_C(0x749549ab)
4114	# define CX8_MISMATCH_LO UINT32_C(0x90f18981)
4115	# define CX8_MISMATCH_HI UINT32_C(0xfd5b4000)
4116	# define CX8_STORE_LO UINT32_C(0x51f6559b)
4117	# define CX8_STORE_HI UINT32_C(0xd1b54963)
4118
4119	PRTUINT64U pBuf = (PRTUINT64U)&pau64->au8[offBuf];
4120
4121	ExpectCtx.rax.u = Ctx.rax.u = CX8_MISMATCH_LO;
4122	ExpectCtx.rdx.u = Ctx.rdx.u = CX8_MISMATCH_HI;
4123
4124	Bs3RegCtxSetGrpSegFromCurPtr(&Ctx, &Ctx.rdi, &Ctx.fs, pBuf);
4125	Bs3RegCtxSetGrpSegFromCurPtr(&ExpectCtx, &ExpectCtx.rdi, &ExpectCtx.fs, pBuf);
4126
4127	for (iMatch = 0; iMatch < 2; iMatch++)
4128	{
4129	uint8_t bExpectXcpt;
4130	pBuf->s.Lo = CX8_OLD_LO;
4131	pBuf->s.Hi = CX8_OLD_HI;
4132
4133	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
4134	g_usBs3TestStep++;
4135	//Bs3TestPrintf("Test: iFlags=%d offBuf=%d iMatch=%u iWorker=%u\n", iFlags, offBuf, iMatch, iWorker);
4136	bExpectXcpt = X86_XCPT_UD;
4137	if (fSupportCX8)
4138	{
4139	if ( offBuf
4140	&& bRing == 3
4141	&& bMode != BS3_MODE_RM
4142	&& !BS3_MODE_IS_V86(bMode)
4143	&& iCfg)
4144	{
4145	bExpectXcpt = X86_XCPT_AC;
4146	ExpectCtx.rflags.u32 = Ctx.rflags.u32;
4147	}
4148	else
4149	{
4150	bExpectXcpt = X86_XCPT_DB;
4151
4152	ExpectCtx.rax.u = CX8_OLD_LO;
4153	ExpectCtx.rdx.u = CX8_OLD_HI;
4154
4155	if (iMatch & 1)
4156	ExpectCtx.rflags.u32 = Ctx.rflags.u32 \| X86_EFL_ZF;
4157	else
4158	ExpectCtx.rflags.u32 = Ctx.rflags.u32 & ~X86_EFL_ZF;
4159	}
4160
4161	/* Kludge! Looks like EFLAGS.AC is cleared when raising #GP in real mode on an i7-6700K. WEIRD! */
4162	if (bMode == BS3_MODE_RM && (Ctx.rflags.u32 & X86_EFL_AC))
4163	{
4164	if (TrapFrame.Ctx.rflags.u32 & X86_EFL_AC)
4165	Bs3TestFailedF("Expected EFLAGS.AC to be cleared (bXcpt=%d)", TrapFrame.bXcpt);
4166	TrapFrame.Ctx.rflags.u32 \|= X86_EFL_AC;
4167	}
4168	}
4169
4170	if ( !Bs3TestCheckRegCtxEx(&TrapFrame.Ctx, &ExpectCtx, bExpectXcpt == X86_XCPT_DB ? s_aWorkers[iWorker].offIcebp : 0, 0 /cbSpAdjust/,
4171	bExpectXcpt == X86_XCPT_DB \|\| BS3_MODE_IS_16BIT_SYS(bMode) ? 0 : X86_EFL_RF, pszMode, g_usBs3TestStep)
4172	\|\| TrapFrame.bXcpt != bExpectXcpt)
4173	{
4174	if (TrapFrame.bXcpt != bExpectXcpt)
4175	Bs3TestFailedF("Expected bXcpt=#%x, got %#x (%#x)", bExpectXcpt, TrapFrame.bXcpt, TrapFrame.uErrCd);
4176	Bs3TestFailedF("^^^ bRing=%u iCfg=%d iWorker=%d iFlags=%d offBuf=%d iMatch=%u\n",
4177	bRing, iCfg, iWorker, iFlags, offBuf, iMatch);
4178	ASMHalt();
4179	}
4180
4181	ExpectCtx.rax.u = Ctx.rax.u = CX8_OLD_LO;
4182	ExpectCtx.rdx.u = Ctx.rdx.u = CX8_OLD_HI;
4183	}
4184	}
4185	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
4186	}
4187	}
4188	} /* for each test config. */
4189
4190	/*
4191	* Next ring.
4192	*/
4193	bRing++;
4194	if (bRing > 3 \|\| bMode == BS3_MODE_RM)
4195	break;
4196	Bs3RegCtxConvertToRingX(&Ctx, bRing);
4197	}
4198
4199	return 0;
4200	}
4201
4202
4203
4204	/*
4205	*
4206	*/
4207	# if ARCH_BITS == 64
4208
4209	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_cmpxchg16b)(uint8_t bMode)
4210	{
4211	BS3REGCTX Ctx;
4212	BS3REGCTX ExpectCtx;
4213	BS3TRAPFRAME TrapFrame;
4214	RTUINT128U au128[3];
4215	PRTUINT128U pau128 = RT_ALIGN_PT(&au128[0], sizeof(RTUINT128U), PRTUINT128U);
4216	bool const fSupportCX16 = RT_BOOL(ASMCpuId_ECX(1) & X86_CPUID_FEATURE_ECX_CX16);
4217	const char BS3_FAR * const pszMode = Bs3GetModeName(bMode);
4218	uint8_t bRing = BS3_MODE_IS_V86(bMode) ? 3 : 0;
4219	unsigned iFlags;
4220	unsigned offBuf;
4221	unsigned iMatch;
4222	unsigned iWorker;
4223	static struct
4224	{
4225	bool fLocked;
4226	uint8_t offUd2;
4227	FNBS3FAR *pfnWorker;
4228	} const s_aWorkers[] =
4229	{
4230	{ false, 4, BS3_CMN_NM(bs3CpuInstr2_cmpxchg16b_rdi_ud2) },
4231	{ false, 5, BS3_CMN_NM(bs3CpuInstr2_o16_cmpxchg16b_rdi_ud2) },
4232	{ false, 5, BS3_CMN_NM(bs3CpuInstr2_repz_cmpxchg16b_rdi_ud2) },
4233	{ false, 5, BS3_CMN_NM(bs3CpuInstr2_repnz_cmpxchg16b_rdi_ud2) },
4234	{ true, 1+4, BS3_CMN_NM(bs3CpuInstr2_lock_cmpxchg16b_rdi_ud2) },
4235	{ true, 1+5, BS3_CMN_NM(bs3CpuInstr2_lock_o16_cmpxchg16b_rdi_ud2) },
4236	{ true, 1+5, BS3_CMN_NM(bs3CpuInstr2_lock_repz_cmpxchg16b_rdi_ud2) },
4237	{ true, 1+5, BS3_CMN_NM(bs3CpuInstr2_lock_repnz_cmpxchg16b_rdi_ud2) },
4238	};
4239
4240	/* Ensure the structures are allocated before we sample the stack pointer. */
4241	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
4242	Bs3MemSet(&ExpectCtx, 0, sizeof(ExpectCtx));
4243	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
4244	Bs3MemSet(pau128, 0, sizeof(pau128[0]) * 2);
4245
4246	/*
4247	* Create test context.
4248	*/
4249	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
4250	if (!fSupportCX16)
4251	Bs3TestPrintf("Note! CMPXCHG16B is not supported by the CPU!\n");
4252
4253
4254	/*
4255	* Run the tests in all rings since alignment issues may behave
4256	* differently in ring-3 compared to ring-0.
4257	*/
4258	for (;;)
4259	{
4260	/*
4261	* One loop with alignment checks disabled and one with enabled.
4262	*/
4263	unsigned iCfg;
4264	for (iCfg = 0; iCfg < 2; iCfg++)
4265	{
4266	if (iCfg)
4267	{
4268	Ctx.rflags.u32 \|= X86_EFL_AC;
4269	Ctx.cr0.u32 \|= X86_CR0_AM;
4270	}
4271	else
4272	{
4273	Ctx.rflags.u32 &= ~X86_EFL_AC;
4274	Ctx.cr0.u32 &= ~X86_CR0_AM;
4275	}
4276
4277	/*
4278	* One loop with the normal variant and one with the locked one
4279	*/
4280	g_usBs3TestStep = 0;
4281	for (iWorker = 0; iWorker < RT_ELEMENTS(s_aWorkers); iWorker++)
4282	{
4283	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, s_aWorkers[iWorker].pfnWorker);
4284
4285	/*
4286	* One loop with all status flags set, and one with them clear.
4287	*/
4288	Ctx.rflags.u16 \|= X86_EFL_STATUS_BITS;
4289	for (iFlags = 0; iFlags < 2; iFlags++)
4290	{
4291	Bs3MemCpy(&ExpectCtx, &Ctx, sizeof(ExpectCtx));
4292
4293	for (offBuf = 0; offBuf < sizeof(RTUINT128U); offBuf++)
4294	{
4295	# define CX16_OLD_LO UINT64_C(0xabb6345dcc9c4bbd)
4296	# define CX16_OLD_HI UINT64_C(0x7b06ea35749549ab)
4297	# define CX16_MISMATCH_LO UINT64_C(0xbace3e3590f18981)
4298	# define CX16_MISMATCH_HI UINT64_C(0x9b385e8bfd5b4000)
4299	# define CX16_STORE_LO UINT64_C(0x5cbd27d251f6559b)
4300	# define CX16_STORE_HI UINT64_C(0x17ff434ed1b54963)
4301
4302	PRTUINT128U pBuf = (PRTUINT128U)&pau128->au8[offBuf];
4303
4304	ExpectCtx.rax.u = Ctx.rax.u = CX16_MISMATCH_LO;
4305	ExpectCtx.rdx.u = Ctx.rdx.u = CX16_MISMATCH_HI;
4306	for (iMatch = 0; iMatch < 2; iMatch++)
4307	{
4308	uint8_t bExpectXcpt;
4309	pBuf->s.Lo = CX16_OLD_LO;
4310	pBuf->s.Hi = CX16_OLD_HI;
4311	ExpectCtx.rdi.u = Ctx.rdi.u = (uintptr_t)pBuf;
4312	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
4313	g_usBs3TestStep++;
4314	//Bs3TestPrintf("Test: iFlags=%d offBuf=%d iMatch=%u iWorker=%u\n", iFlags, offBuf, iMatch, iWorker);
4315	bExpectXcpt = X86_XCPT_UD;
4316	if (fSupportCX16)
4317	{
4318	if (offBuf & 15)
4319	{
4320	bExpectXcpt = X86_XCPT_GP;
4321	ExpectCtx.rip.u = Ctx.rip.u;
4322	ExpectCtx.rflags.u32 = Ctx.rflags.u32;
4323	}
4324	else
4325	{
4326	ExpectCtx.rax.u = CX16_OLD_LO;
4327	ExpectCtx.rdx.u = CX16_OLD_HI;
4328	if (iMatch & 1)
4329	ExpectCtx.rflags.u32 = Ctx.rflags.u32 \| X86_EFL_ZF;
4330	else
4331	ExpectCtx.rflags.u32 = Ctx.rflags.u32 & ~X86_EFL_ZF;
4332	ExpectCtx.rip.u = Ctx.rip.u + s_aWorkers[iWorker].offUd2;
4333	}
4334	ExpectCtx.rflags.u32 \|= X86_EFL_RF;
4335	}
4336	if ( !Bs3TestCheckRegCtxEx(&TrapFrame.Ctx, &ExpectCtx, 0 /cbPcAdjust/, 0 /cbSpAdjust/,
4337	0 /fExtraEfl/, pszMode, 0 /idTestStep/)
4338	\|\| TrapFrame.bXcpt != bExpectXcpt)
4339	{
4340	if (TrapFrame.bXcpt != bExpectXcpt)
4341	Bs3TestFailedF("Expected bXcpt=#%x, got %#x (%#x)", bExpectXcpt, TrapFrame.bXcpt, TrapFrame.uErrCd);
4342	Bs3TestFailedF("^^^bRing=%d iWorker=%d iFlags=%d offBuf=%d iMatch=%u\n", bRing, iWorker, iFlags, offBuf, iMatch);
4343	ASMHalt();
4344	}
4345
4346	ExpectCtx.rax.u = Ctx.rax.u = CX16_OLD_LO;
4347	ExpectCtx.rdx.u = Ctx.rdx.u = CX16_OLD_HI;
4348	}
4349	}
4350	Ctx.rflags.u16 &= ~X86_EFL_STATUS_BITS;
4351	}
4352	}
4353	} /* for each test config. */
4354
4355	/*
4356	* Next ring.
4357	*/
4358	bRing++;
4359	if (bRing > 3 \|\| bMode == BS3_MODE_RM)
4360	break;
4361	Bs3RegCtxConvertToRingX(&Ctx, bRing);
4362	}
4363
4364	return 0;
4365	}
4366
4367
4368	static void bs3CpuInstr2_fsgsbase_ExpectUD(uint8_t bMode, PBS3REGCTX pCtx, PBS3REGCTX pExpectCtx, PBS3TRAPFRAME pTrapFrame)
4369	{
4370	pCtx->rbx.u = 0;
4371	Bs3MemCpy(pExpectCtx, pCtx, sizeof(*pExpectCtx));
4372	Bs3TrapSetJmpAndRestore(pCtx, pTrapFrame);
4373	pExpectCtx->rip.u = pCtx->rip.u;
4374	pExpectCtx->rflags.u32 \|= X86_EFL_RF;
4375	if ( !Bs3TestCheckRegCtxEx(&pTrapFrame->Ctx, pExpectCtx, 0 /cbPcAdjust/, 0 /cbSpAdjust/, 0 /fExtraEfl/, "lm64",
4376	0 /idTestStep/)
4377	\|\| pTrapFrame->bXcpt != X86_XCPT_UD)
4378	{
4379	Bs3TestFailedF("Expected #UD, got %#x (%#x)", pTrapFrame->bXcpt, pTrapFrame->uErrCd);
4380	ASMHalt();
4381	}
4382	}
4383
4384
4385	static bool bs3CpuInstr2_fsgsbase_VerifyWorker(uint8_t bMode, PBS3REGCTX pCtx, PBS3REGCTX pExpectCtx, PBS3TRAPFRAME pTrapFrame,
4386	BS3CI2FSGSBASE const pFsGsBaseWorker, unsigned puIter)
4387	{
4388	bool fPassed = true;
4389	unsigned iValue = 0;
4390	static const struct
4391	{
4392	bool fGP;
4393	uint64_t u64Base;
4394	} s_aValues64[] =
4395	{
4396	{ false, UINT64_C(0x0000000000000000) },
4397	{ false, UINT64_C(0x0000000000000001) },
4398	{ false, UINT64_C(0x0000000000000010) },
4399	{ false, UINT64_C(0x0000000000000123) },
4400	{ false, UINT64_C(0x0000000000001234) },
4401	{ false, UINT64_C(0x0000000000012345) },
4402	{ false, UINT64_C(0x0000000000123456) },
4403	{ false, UINT64_C(0x0000000001234567) },
4404	{ false, UINT64_C(0x0000000012345678) },
4405	{ false, UINT64_C(0x0000000123456789) },
4406	{ false, UINT64_C(0x000000123456789a) },
4407	{ false, UINT64_C(0x00000123456789ab) },
4408	{ false, UINT64_C(0x0000123456789abc) },
4409	{ false, UINT64_C(0x00007ffffeefefef) },
4410	{ false, UINT64_C(0x00007fffffffffff) },
4411	{ true, UINT64_C(0x0000800000000000) },
4412	{ true, UINT64_C(0x0000800000000000) },
4413	{ true, UINT64_C(0x0000800000000333) },
4414	{ true, UINT64_C(0x0001000000000000) },
4415	{ true, UINT64_C(0x0012000000000000) },
4416	{ true, UINT64_C(0x0123000000000000) },
4417	{ true, UINT64_C(0x1234000000000000) },
4418	{ true, UINT64_C(0xffff300000000000) },
4419	{ true, UINT64_C(0xffff7fffffffffff) },
4420	{ true, UINT64_C(0xffff7fffffffffff) },
4421	{ false, UINT64_C(0xffff800000000000) },
4422	{ false, UINT64_C(0xffffffffffeefefe) },
4423	{ false, UINT64_C(0xffffffffffffffff) },
4424	{ false, UINT64_C(0xffffffffffffffff) },
4425	{ false, UINT64_C(0x00000000efefefef) },
4426	{ false, UINT64_C(0x0000000080204060) },
4427	{ false, UINT64_C(0x00000000ddeeffaa) },
4428	{ false, UINT64_C(0x00000000fdecdbca) },
4429	{ false, UINT64_C(0x000000006098456b) },
4430	{ false, UINT64_C(0x0000000098506099) },
4431	{ false, UINT64_C(0x00000000206950bc) },
4432	{ false, UINT64_C(0x000000009740395d) },
4433	{ false, UINT64_C(0x0000000064a9455e) },
4434	{ false, UINT64_C(0x00000000d20b6eff) },
4435	{ false, UINT64_C(0x0000000085296d46) },
4436	{ false, UINT64_C(0x0000000007000039) },
4437	{ false, UINT64_C(0x000000000007fe00) },
4438	};
4439
4440	Bs3RegCtxSetRipCsFromCurPtr(pCtx, pFsGsBaseWorker->pfnVerifyWorker);
4441	if (pFsGsBaseWorker->f64BitOperand)
4442	{
4443	for (iValue = 0; iValue < RT_ELEMENTS(s_aValues64); iValue++)
4444	{
4445	bool const fGP = s_aValues64[iValue].fGP;
4446
4447	pCtx->rbx.u = s_aValues64[iValue].u64Base;
4448	pCtx->rcx.u = 0;
4449	pCtx->cr4.u \|= X86_CR4_FSGSBASE;
4450	Bs3MemCpy(pExpectCtx, pCtx, sizeof(*pExpectCtx));
4451	Bs3TrapSetJmpAndRestore(pCtx, pTrapFrame);
4452	pExpectCtx->rip.u = pCtx->rip.u + (!fGP ? pFsGsBaseWorker->offVerifyWorkerUd2 : 0);
4453	pExpectCtx->rbx.u = !fGP ? 0 : s_aValues64[iValue].u64Base;
4454	pExpectCtx->rcx.u = !fGP ? s_aValues64[iValue].u64Base : 0;
4455	pExpectCtx->rflags.u32 \|= X86_EFL_RF;
4456	if ( !Bs3TestCheckRegCtxEx(&pTrapFrame->Ctx, pExpectCtx, 0 /cbPcAdjust/, 0 /cbSpAdjust/,
4457	0 /fExtraEfl/, "lm64", 0 /idTestStep/)
4458	\|\| (fGP && pTrapFrame->bXcpt != X86_XCPT_GP))
4459	{
4460	if (fGP && pTrapFrame->bXcpt != X86_XCPT_GP)
4461	Bs3TestFailedF("Expected #GP, got %#x (%#x)", pTrapFrame->bXcpt, pTrapFrame->uErrCd);
4462	else
4463	Bs3TestFailedF("iValue=%u\n", iValue);
4464	fPassed = false;
4465	break;
4466	}
4467	}
4468	}
4469	else
4470	{
4471	for (iValue = 0; iValue < RT_ELEMENTS(s_aValues64); iValue++)
4472	{
4473	pCtx->rbx.u = s_aValues64[iValue].u64Base;
4474	pCtx->rcx.u = ~s_aValues64[iValue].u64Base;
4475	pCtx->cr4.u \|= X86_CR4_FSGSBASE;
4476	Bs3MemCpy(pExpectCtx, pCtx, sizeof(*pExpectCtx));
4477	Bs3TrapSetJmpAndRestore(pCtx, pTrapFrame);
4478	pExpectCtx->rip.u = pCtx->rip.u + pFsGsBaseWorker->offVerifyWorkerUd2;
4479	pExpectCtx->rbx.u = 0;
4480	pExpectCtx->rcx.u = s_aValues64[iValue].u64Base & UINT64_C(0x00000000ffffffff);
4481	pExpectCtx->rflags.u32 \|= X86_EFL_RF;
4482	if (!Bs3TestCheckRegCtxEx(&pTrapFrame->Ctx, pExpectCtx, 0 /cbPcAdjust/, 0 /cbSpAdjust/,
4483	0 /fExtraEfl/, "lm64", 0 /idTestStep/))
4484	{
4485	Bs3TestFailedF("iValue=%u\n", iValue);
4486	fPassed = false;
4487	break;
4488	}
4489	}
4490	}
4491
4492	*puIter = iValue;
4493	return fPassed;
4494	}
4495
4496
4497	static void bs3CpuInstr2_rdfsbase_rdgsbase_Common(uint8_t bMode, BS3CI2FSGSBASE const *paFsGsBaseWorkers,
4498	unsigned cFsGsBaseWorkers, uint32_t idxFsGsBaseMsr)
4499	{
4500	BS3REGCTX Ctx;
4501	BS3REGCTX ExpectCtx;
4502	BS3TRAPFRAME TrapFrame;
4503	unsigned iWorker;
4504	unsigned iIter;
4505	uint32_t uDummy;
4506	uint32_t uStdExtFeatEbx;
4507	bool fSupportsFsGsBase;
4508
4509	ASMCpuId_Idx_ECX(7, 0, &uDummy, &uStdExtFeatEbx, &uDummy, &uDummy);
4510	fSupportsFsGsBase = RT_BOOL(uStdExtFeatEbx & X86_CPUID_STEXT_FEATURE_EBX_FSGSBASE);
4511
4512	/* Ensure the structures are allocated before we sample the stack pointer. */
4513	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
4514	Bs3MemSet(&ExpectCtx, 0, sizeof(ExpectCtx));
4515	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
4516
4517	/*
4518	* Create test context.
4519	*/
4520	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
4521
4522	for (iWorker = 0; iWorker < cFsGsBaseWorkers; iWorker++)
4523	{
4524	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, paFsGsBaseWorkers[iWorker].pfnWorker);
4525	if (fSupportsFsGsBase)
4526	{
4527	uint64_t const uBaseAddr = ASMRdMsr(idxFsGsBaseMsr);
4528
4529	/* CR4.FSGSBASE disabled -> #UD. */
4530	Ctx.cr4.u &= ~X86_CR4_FSGSBASE;
4531	bs3CpuInstr2_fsgsbase_ExpectUD(bMode, &Ctx, &ExpectCtx, &TrapFrame);
4532
4533	/* Read and verify existing base address. */
4534	Ctx.rbx.u = 0;
4535	Ctx.cr4.u \|= X86_CR4_FSGSBASE;
4536	Bs3MemCpy(&ExpectCtx, &Ctx, sizeof(ExpectCtx));
4537	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
4538	ExpectCtx.rip.u = Ctx.rip.u + paFsGsBaseWorkers[iWorker].offWorkerUd2;
4539	ExpectCtx.rbx.u = uBaseAddr;
4540	ExpectCtx.rflags.u32 \|= X86_EFL_RF;
4541	if (!Bs3TestCheckRegCtxEx(&TrapFrame.Ctx, &ExpectCtx, 0 /cbPcAdjust/, 0 /cbSpAdjust/, 0 /fExtraEfl/, "lm64",
4542	0 /idTestStep/))
4543	{
4544	ASMHalt();
4545	}
4546
4547	/* Write, read and verify series of base addresses. */
4548	if (!bs3CpuInstr2_fsgsbase_VerifyWorker(bMode, &Ctx, &ExpectCtx, &TrapFrame, &paFsGsBaseWorkers[iWorker], &iIter))
4549	{
4550	Bs3TestFailedF("^^^ %s: iWorker=%u iIter=%u\n", paFsGsBaseWorkers[iWorker].pszDesc, iWorker, iIter);
4551	ASMHalt();
4552	}
4553
4554	/* Restore original base address. */
4555	ASMWrMsr(idxFsGsBaseMsr, uBaseAddr);
4556
4557	/* Clean used GPRs. */
4558	Ctx.rbx.u = 0;
4559	Ctx.rcx.u = 0;
4560	}
4561	else
4562	{
4563	/* Unsupported by CPUID -> #UD. */
4564	Bs3TestPrintf("Note! FSGSBASE is not supported by the CPU!\n");
4565	bs3CpuInstr2_fsgsbase_ExpectUD(bMode, &Ctx, &ExpectCtx, &TrapFrame);
4566	}
4567	}
4568	}
4569
4570
4571	static void bs3CpuInstr2_wrfsbase_wrgsbase_Common(uint8_t bMode, BS3CI2FSGSBASE const *paFsGsBaseWorkers,
4572	unsigned cFsGsBaseWorkers, uint32_t idxFsGsBaseMsr)
4573	{
4574	BS3REGCTX Ctx;
4575	BS3REGCTX ExpectCtx;
4576	BS3TRAPFRAME TrapFrame;
4577	unsigned iWorker;
4578	unsigned iIter;
4579	uint32_t uDummy;
4580	uint32_t uStdExtFeatEbx;
4581	bool fSupportsFsGsBase;
4582
4583	ASMCpuId_Idx_ECX(7, 0, &uDummy, &uStdExtFeatEbx, &uDummy, &uDummy);
4584	fSupportsFsGsBase = RT_BOOL(uStdExtFeatEbx & X86_CPUID_STEXT_FEATURE_EBX_FSGSBASE);
4585
4586	/* Ensure the structures are allocated before we sample the stack pointer. */
4587	Bs3MemSet(&Ctx, 0, sizeof(Ctx));
4588	Bs3MemSet(&ExpectCtx, 0, sizeof(ExpectCtx));
4589	Bs3MemSet(&TrapFrame, 0, sizeof(TrapFrame));
4590
4591	/*
4592	* Create test context.
4593	*/
4594	Bs3RegCtxSaveEx(&Ctx, bMode, 512);
4595
4596	for (iWorker = 0; iWorker < cFsGsBaseWorkers; iWorker++)
4597	{
4598	Bs3RegCtxSetRipCsFromCurPtr(&Ctx, paFsGsBaseWorkers[iWorker].pfnWorker);
4599	if (fSupportsFsGsBase)
4600	{
4601	uint64_t const uBaseAddr = ASMRdMsr(idxFsGsBaseMsr);
4602
4603	/* CR4.FSGSBASE disabled -> #UD. */
4604	Ctx.cr4.u &= ~X86_CR4_FSGSBASE;
4605	bs3CpuInstr2_fsgsbase_ExpectUD(bMode, &Ctx, &ExpectCtx, &TrapFrame);
4606
4607	/* Write a base address. */
4608	Ctx.rbx.u = 0xa0000;
4609	Ctx.cr4.u \|= X86_CR4_FSGSBASE;
4610	Bs3MemCpy(&ExpectCtx, &Ctx, sizeof(ExpectCtx));
4611	Bs3TrapSetJmpAndRestore(&Ctx, &TrapFrame);
4612	ExpectCtx.rip.u = Ctx.rip.u + paFsGsBaseWorkers[iWorker].offWorkerUd2;
4613	ExpectCtx.rflags.u32 \|= X86_EFL_RF;
4614	if (!Bs3TestCheckRegCtxEx(&TrapFrame.Ctx, &ExpectCtx, 0 /cbPcAdjust/, 0 /cbSpAdjust/, 0 /fExtraEfl/, "lm64",
4615	0 /idTestStep/))
4616	{
4617	ASMHalt();
4618	}
4619
4620	/* Write and read back series of base addresses. */
4621	if (!bs3CpuInstr2_fsgsbase_VerifyWorker(bMode, &Ctx, &ExpectCtx, &TrapFrame, &paFsGsBaseWorkers[iWorker], &iIter))
4622	{
4623	Bs3TestFailedF("^^^ %s: iWorker=%u iIter=%u\n", paFsGsBaseWorkers[iWorker].pszDesc, iWorker, iIter);
4624	ASMHalt();
4625	}
4626
4627	/* Restore original base address. */
4628	ASMWrMsr(idxFsGsBaseMsr, uBaseAddr);
4629
4630	/* Clean used GPRs. */
4631	Ctx.rbx.u = 0;
4632	Ctx.rcx.u = 0;
4633	}
4634	else
4635	{
4636	/* Unsupported by CPUID -> #UD. */
4637	Bs3TestPrintf("Note! FSGSBASE is not supported by the CPU!\n");
4638	bs3CpuInstr2_fsgsbase_ExpectUD(bMode, &Ctx, &ExpectCtx, &TrapFrame);
4639	}
4640	}
4641	}
4642
4643
4644	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_wrfsbase)(uint8_t bMode)
4645	{
4646	bs3CpuInstr2_wrfsbase_wrgsbase_Common(bMode, s_aWrFsBaseWorkers, RT_ELEMENTS(s_aWrFsBaseWorkers), MSR_K8_FS_BASE);
4647	return 0;
4648	}
4649
4650
4651	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_wrgsbase)(uint8_t bMode)
4652	{
4653	bs3CpuInstr2_wrfsbase_wrgsbase_Common(bMode, s_aWrGsBaseWorkers, RT_ELEMENTS(s_aWrGsBaseWorkers), MSR_K8_GS_BASE);
4654	return 0;
4655	}
4656
4657
4658	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_rdfsbase)(uint8_t bMode)
4659	{
4660	bs3CpuInstr2_rdfsbase_rdgsbase_Common(bMode, s_aRdFsBaseWorkers, RT_ELEMENTS(s_aRdFsBaseWorkers), MSR_K8_FS_BASE);
4661	return 0;
4662	}
4663
4664
4665	BS3_DECL_FAR(uint8_t) BS3_CMN_NM(bs3CpuInstr2_rdgsbase)(uint8_t bMode)
4666	{
4667	bs3CpuInstr2_rdfsbase_rdgsbase_Common(bMode, s_aRdGsBaseWorkers, RT_ELEMENTS(s_aRdGsBaseWorkers), MSR_K8_GS_BASE);
4668	return 0;
4669	}
4670
4671	# endif /* ARCH_BITS == 64 */
4672
4673	#endif /* BS3_INSTANTIATING_CMN */
4674
4675
4676
4677	/*
4678	* Mode specific code.
4679	* Mode specific code.
4680	* Mode specific code.
4681	*/
4682	#ifdef BS3_INSTANTIATING_MODE
4683
4684
4685	#endif /* BS3_INSTANTIATING_MODE */
4686

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

source: vbox/trunk/src/VBox/ValidationKit/bootsectors/bs3-cpu-instr-2-template.c@ 105686

Download in other formats: