test-i386.c@ 77978

Last change on this file since 77978 was 69465, checked in by vboxsync, 7 years ago
recompiler: scm updates
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 70.5 KB

Line
1	/*
2	* x86 CPU test
3	*
4	* Copyright (c) 2003 Fabrice Bellard
5	*
6	* This program is free software; you can redistribute it and/or modify
7	* it under the terms of the GNU General Public License as published by
8	* the Free Software Foundation; either version 2 of the License, or
9	* (at your option) any later version.
10	*
11	* This program is distributed in the hope that it will be useful,
12	* but WITHOUT ANY WARRANTY; without even the implied warranty of
13	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14	* GNU General Public License for more details.
15	*
16	* You should have received a copy of the GNU General Public License
17	* along with this program; if not, see <http://www.gnu.org/licenses/>.
18	*/
19
20	/*
21	* Oracle GPL Disclaimer: For the avoidance of doubt, except that if any license choice
22	* other than GPL or LGPL is available it will apply instead, Oracle elects to use only
23	* the General Public License version 2 (GPLv2) at this time for any software where
24	* a choice of GPL license versions is made available with the language indicating
25	* that GPLv2 or any later version may be used, or where a choice of which version
26	* of the GPL is applied is otherwise unspecified.
27	*/
28
29	#define _GNU_SOURCE
30	#include <stdlib.h>
31	#include <stdio.h>
32	#include <string.h>
33	#include <inttypes.h>
34	#include <math.h>
35	#include <signal.h>
36	#include <setjmp.h>
37	#include <errno.h>
38	#include <sys/ucontext.h>
39	#include <sys/mman.h>
40
41	#if !defined(__x86_64__)
42	//#define TEST_VM86
43	#define TEST_SEGS
44	#endif
45	//#define LINUX_VM86_IOPL_FIX
46	//#define TEST_P4_FLAGS
47	#ifdef __SSE__
48	#define TEST_SSE
49	#define TEST_CMOV 1
50	#define TEST_FCOMI 1
51	#else
52	#undef TEST_SSE
53	#define TEST_CMOV 1
54	#define TEST_FCOMI 1
55	#endif
56
57	#if defined(__x86_64__)
58	#define FMT64X "%016lx"
59	#define FMTLX "%016lx"
60	#define X86_64_ONLY(x) x
61	#else
62	#define FMT64X "%016" PRIx64
63	#define FMTLX "%08lx"
64	#define X86_64_ONLY(x)
65	#endif
66
67	#ifdef TEST_VM86
68	#include <asm/vm86.h>
69	#endif
70
71	#define xglue(x, y) x ## y
72	#define glue(x, y) xglue(x, y)
73	#define stringify(s) tostring(s)
74	#define tostring(s) #s
75
76	#define CC_C 0x0001
77	#define CC_P 0x0004
78	#define CC_A 0x0010
79	#define CC_Z 0x0040
80	#define CC_S 0x0080
81	#define CC_O 0x0800
82
83	#define __init_call __attribute__ ((unused,__section__ ("initcall")))
84
85	#define CC_MASK (CC_C \| CC_P \| CC_Z \| CC_S \| CC_O \| CC_A)
86
87	#if defined(__x86_64__)
88	static inline long i2l(long v)
89	{
90	return v \| ((v ^ 0xabcd) << 32);
91	}
92	#else
93	static inline long i2l(long v)
94	{
95	return v;
96	}
97	#endif
98
99	#define OP add
100	#include "test-i386.h"
101
102	#define OP sub
103	#include "test-i386.h"
104
105	#define OP xor
106	#include "test-i386.h"
107
108	#define OP and
109	#include "test-i386.h"
110
111	#define OP or
112	#include "test-i386.h"
113
114	#define OP cmp
115	#include "test-i386.h"
116
117	#define OP adc
118	#define OP_CC
119	#include "test-i386.h"
120
121	#define OP sbb
122	#define OP_CC
123	#include "test-i386.h"
124
125	#define OP inc
126	#define OP_CC
127	#define OP1
128	#include "test-i386.h"
129
130	#define OP dec
131	#define OP_CC
132	#define OP1
133	#include "test-i386.h"
134
135	#define OP neg
136	#define OP_CC
137	#define OP1
138	#include "test-i386.h"
139
140	#define OP not
141	#define OP_CC
142	#define OP1
143	#include "test-i386.h"
144
145	#undef CC_MASK
146	#define CC_MASK (CC_C \| CC_P \| CC_Z \| CC_S \| CC_O)
147
148	#define OP shl
149	#include "test-i386-shift.h"
150
151	#define OP shr
152	#include "test-i386-shift.h"
153
154	#define OP sar
155	#include "test-i386-shift.h"
156
157	#define OP rol
158	#include "test-i386-shift.h"
159
160	#define OP ror
161	#include "test-i386-shift.h"
162
163	#define OP rcr
164	#define OP_CC
165	#include "test-i386-shift.h"
166
167	#define OP rcl
168	#define OP_CC
169	#include "test-i386-shift.h"
170
171	#define OP shld
172	#define OP_SHIFTD
173	#define OP_NOBYTE
174	#include "test-i386-shift.h"
175
176	#define OP shrd
177	#define OP_SHIFTD
178	#define OP_NOBYTE
179	#include "test-i386-shift.h"
180
181	/* XXX: should be more precise ? */
182	#undef CC_MASK
183	#define CC_MASK (CC_C)
184
185	#define OP bt
186	#define OP_NOBYTE
187	#include "test-i386-shift.h"
188
189	#define OP bts
190	#define OP_NOBYTE
191	#include "test-i386-shift.h"
192
193	#define OP btr
194	#define OP_NOBYTE
195	#include "test-i386-shift.h"
196
197	#define OP btc
198	#define OP_NOBYTE
199	#include "test-i386-shift.h"
200
201	/* lea test (modrm support) */
202	#define TEST_LEAQ(STR)\
203	{\
204	asm("lea " STR ", %0"\
205	: "=r" (res)\
206	: "a" (eax), "b" (ebx), "c" (ecx), "d" (edx), "S" (esi), "D" (edi));\
207	printf("lea %s = " FMTLX "\n", STR, res);\
208	}
209
210	#define TEST_LEA(STR)\
211	{\
212	asm("lea " STR ", %0"\
213	: "=r" (res)\
214	: "a" (eax), "b" (ebx), "c" (ecx), "d" (edx), "S" (esi), "D" (edi));\
215	printf("lea %s = " FMTLX "\n", STR, res);\
216	}
217
218	#define TEST_LEA16(STR)\
219	{\
220	asm(".code16 ; .byte 0x67 ; leal " STR ", %0 ; .code32"\
221	: "=wq" (res)\
222	: "a" (eax), "b" (ebx), "c" (ecx), "d" (edx), "S" (esi), "D" (edi));\
223	printf("lea %s = %08lx\n", STR, res);\
224	}
225
226
227	void test_lea(void)
228	{
229	long eax, ebx, ecx, edx, esi, edi, res;
230	eax = i2l(0x0001);
231	ebx = i2l(0x0002);
232	ecx = i2l(0x0004);
233	edx = i2l(0x0008);
234	esi = i2l(0x0010);
235	edi = i2l(0x0020);
236
237	TEST_LEA("0x4000");
238
239	TEST_LEA("(%%eax)");
240	TEST_LEA("(%%ebx)");
241	TEST_LEA("(%%ecx)");
242	TEST_LEA("(%%edx)");
243	TEST_LEA("(%%esi)");
244	TEST_LEA("(%%edi)");
245
246	TEST_LEA("0x40(%%eax)");
247	TEST_LEA("0x40(%%ebx)");
248	TEST_LEA("0x40(%%ecx)");
249	TEST_LEA("0x40(%%edx)");
250	TEST_LEA("0x40(%%esi)");
251	TEST_LEA("0x40(%%edi)");
252
253	TEST_LEA("0x4000(%%eax)");
254	TEST_LEA("0x4000(%%ebx)");
255	TEST_LEA("0x4000(%%ecx)");
256	TEST_LEA("0x4000(%%edx)");
257	TEST_LEA("0x4000(%%esi)");
258	TEST_LEA("0x4000(%%edi)");
259
260	TEST_LEA("(%%eax, %%ecx)");
261	TEST_LEA("(%%ebx, %%edx)");
262	TEST_LEA("(%%ecx, %%ecx)");
263	TEST_LEA("(%%edx, %%ecx)");
264	TEST_LEA("(%%esi, %%ecx)");
265	TEST_LEA("(%%edi, %%ecx)");
266
267	TEST_LEA("0x40(%%eax, %%ecx)");
268	TEST_LEA("0x4000(%%ebx, %%edx)");
269
270	TEST_LEA("(%%ecx, %%ecx, 2)");
271	TEST_LEA("(%%edx, %%ecx, 4)");
272	TEST_LEA("(%%esi, %%ecx, 8)");
273
274	TEST_LEA("(,%%eax, 2)");
275	TEST_LEA("(,%%ebx, 4)");
276	TEST_LEA("(,%%ecx, 8)");
277
278	TEST_LEA("0x40(,%%eax, 2)");
279	TEST_LEA("0x40(,%%ebx, 4)");
280	TEST_LEA("0x40(,%%ecx, 8)");
281
282
283	TEST_LEA("-10(%%ecx, %%ecx, 2)");
284	TEST_LEA("-10(%%edx, %%ecx, 4)");
285	TEST_LEA("-10(%%esi, %%ecx, 8)");
286
287	TEST_LEA("0x4000(%%ecx, %%ecx, 2)");
288	TEST_LEA("0x4000(%%edx, %%ecx, 4)");
289	TEST_LEA("0x4000(%%esi, %%ecx, 8)");
290
291	#if defined(__x86_64__)
292	TEST_LEAQ("0x4000");
293	TEST_LEAQ("0x4000(%%rip)");
294
295	TEST_LEAQ("(%%rax)");
296	TEST_LEAQ("(%%rbx)");
297	TEST_LEAQ("(%%rcx)");
298	TEST_LEAQ("(%%rdx)");
299	TEST_LEAQ("(%%rsi)");
300	TEST_LEAQ("(%%rdi)");
301
302	TEST_LEAQ("0x40(%%rax)");
303	TEST_LEAQ("0x40(%%rbx)");
304	TEST_LEAQ("0x40(%%rcx)");
305	TEST_LEAQ("0x40(%%rdx)");
306	TEST_LEAQ("0x40(%%rsi)");
307	TEST_LEAQ("0x40(%%rdi)");
308
309	TEST_LEAQ("0x4000(%%rax)");
310	TEST_LEAQ("0x4000(%%rbx)");
311	TEST_LEAQ("0x4000(%%rcx)");
312	TEST_LEAQ("0x4000(%%rdx)");
313	TEST_LEAQ("0x4000(%%rsi)");
314	TEST_LEAQ("0x4000(%%rdi)");
315
316	TEST_LEAQ("(%%rax, %%rcx)");
317	TEST_LEAQ("(%%rbx, %%rdx)");
318	TEST_LEAQ("(%%rcx, %%rcx)");
319	TEST_LEAQ("(%%rdx, %%rcx)");
320	TEST_LEAQ("(%%rsi, %%rcx)");
321	TEST_LEAQ("(%%rdi, %%rcx)");
322
323	TEST_LEAQ("0x40(%%rax, %%rcx)");
324	TEST_LEAQ("0x4000(%%rbx, %%rdx)");
325
326	TEST_LEAQ("(%%rcx, %%rcx, 2)");
327	TEST_LEAQ("(%%rdx, %%rcx, 4)");
328	TEST_LEAQ("(%%rsi, %%rcx, 8)");
329
330	TEST_LEAQ("(,%%rax, 2)");
331	TEST_LEAQ("(,%%rbx, 4)");
332	TEST_LEAQ("(,%%rcx, 8)");
333
334	TEST_LEAQ("0x40(,%%rax, 2)");
335	TEST_LEAQ("0x40(,%%rbx, 4)");
336	TEST_LEAQ("0x40(,%%rcx, 8)");
337
338
339	TEST_LEAQ("-10(%%rcx, %%rcx, 2)");
340	TEST_LEAQ("-10(%%rdx, %%rcx, 4)");
341	TEST_LEAQ("-10(%%rsi, %%rcx, 8)");
342
343	TEST_LEAQ("0x4000(%%rcx, %%rcx, 2)");
344	TEST_LEAQ("0x4000(%%rdx, %%rcx, 4)");
345	TEST_LEAQ("0x4000(%%rsi, %%rcx, 8)");
346	#else
347	/* limited 16 bit addressing test */
348	TEST_LEA16("0x4000");
349	TEST_LEA16("(%%bx)");
350	TEST_LEA16("(%%si)");
351	TEST_LEA16("(%%di)");
352	TEST_LEA16("0x40(%%bx)");
353	TEST_LEA16("0x40(%%si)");
354	TEST_LEA16("0x40(%%di)");
355	TEST_LEA16("0x4000(%%bx)");
356	TEST_LEA16("0x4000(%%si)");
357	TEST_LEA16("(%%bx,%%si)");
358	TEST_LEA16("(%%bx,%%di)");
359	TEST_LEA16("0x40(%%bx,%%si)");
360	TEST_LEA16("0x40(%%bx,%%di)");
361	TEST_LEA16("0x4000(%%bx,%%si)");
362	TEST_LEA16("0x4000(%%bx,%%di)");
363	#endif
364	}
365
366	#define TEST_JCC(JCC, v1, v2)\
367	{\
368	int res;\
369	asm("movl $1, %0\n\t"\
370	"cmpl %2, %1\n\t"\
371	"j" JCC " 1f\n\t"\
372	"movl $0, %0\n\t"\
373	"1:\n\t"\
374	: "=r" (res)\
375	: "r" (v1), "r" (v2));\
376	printf("%-10s %d\n", "j" JCC, res);\
377	\
378	asm("movl $0, %0\n\t"\
379	"cmpl %2, %1\n\t"\
380	"set" JCC " %b0\n\t"\
381	: "=r" (res)\
382	: "r" (v1), "r" (v2));\
383	printf("%-10s %d\n", "set" JCC, res);\
384	if (TEST_CMOV) {\
385	long val = i2l(1);\
386	long res = i2l(0x12345678);\
387	X86_64_ONLY(\
388	asm("cmpl %2, %1\n\t"\
389	"cmov" JCC "q %3, %0\n\t"\
390	: "=r" (res)\
391	: "r" (v1), "r" (v2), "m" (val), "0" (res));\
392	printf("%-10s R=" FMTLX "\n", "cmov" JCC "q", res);)\
393	asm("cmpl %2, %1\n\t"\
394	"cmov" JCC "l %k3, %k0\n\t"\
395	: "=r" (res)\
396	: "r" (v1), "r" (v2), "m" (val), "0" (res));\
397	printf("%-10s R=" FMTLX "\n", "cmov" JCC "l", res);\
398	asm("cmpl %2, %1\n\t"\
399	"cmov" JCC "w %w3, %w0\n\t"\
400	: "=r" (res)\
401	: "r" (v1), "r" (v2), "r" (1), "0" (res));\
402	printf("%-10s R=" FMTLX "\n", "cmov" JCC "w", res);\
403	} \
404	}
405
406	/* various jump tests */
407	void test_jcc(void)
408	{
409	TEST_JCC("ne", 1, 1);
410	TEST_JCC("ne", 1, 0);
411
412	TEST_JCC("e", 1, 1);
413	TEST_JCC("e", 1, 0);
414
415	TEST_JCC("l", 1, 1);
416	TEST_JCC("l", 1, 0);
417	TEST_JCC("l", 1, -1);
418
419	TEST_JCC("le", 1, 1);
420	TEST_JCC("le", 1, 0);
421	TEST_JCC("le", 1, -1);
422
423	TEST_JCC("ge", 1, 1);
424	TEST_JCC("ge", 1, 0);
425	TEST_JCC("ge", -1, 1);
426
427	TEST_JCC("g", 1, 1);
428	TEST_JCC("g", 1, 0);
429	TEST_JCC("g", 1, -1);
430
431	TEST_JCC("b", 1, 1);
432	TEST_JCC("b", 1, 0);
433	TEST_JCC("b", 1, -1);
434
435	TEST_JCC("be", 1, 1);
436	TEST_JCC("be", 1, 0);
437	TEST_JCC("be", 1, -1);
438
439	TEST_JCC("ae", 1, 1);
440	TEST_JCC("ae", 1, 0);
441	TEST_JCC("ae", 1, -1);
442
443	TEST_JCC("a", 1, 1);
444	TEST_JCC("a", 1, 0);
445	TEST_JCC("a", 1, -1);
446
447
448	TEST_JCC("p", 1, 1);
449	TEST_JCC("p", 1, 0);
450
451	TEST_JCC("np", 1, 1);
452	TEST_JCC("np", 1, 0);
453
454	TEST_JCC("o", 0x7fffffff, 0);
455	TEST_JCC("o", 0x7fffffff, -1);
456
457	TEST_JCC("no", 0x7fffffff, 0);
458	TEST_JCC("no", 0x7fffffff, -1);
459
460	TEST_JCC("s", 0, 1);
461	TEST_JCC("s", 0, -1);
462	TEST_JCC("s", 0, 0);
463
464	TEST_JCC("ns", 0, 1);
465	TEST_JCC("ns", 0, -1);
466	TEST_JCC("ns", 0, 0);
467	}
468
469	#define TEST_LOOP(insn) \
470	{\
471	for(i = 0; i < sizeof(ecx_vals) / sizeof(long); i++) {\
472	ecx = ecx_vals[i];\
473	for(zf = 0; zf < 2; zf++) {\
474	asm("test %2, %2\n\t"\
475	"movl $1, %0\n\t"\
476	insn " 1f\n\t" \
477	"movl $0, %0\n\t"\
478	"1:\n\t"\
479	: "=a" (res)\
480	: "c" (ecx), "b" (!zf)); \
481	printf("%-10s ECX=" FMTLX " ZF=%ld r=%d\n", insn, ecx, zf, res); \
482	}\
483	}\
484	}
485
486	void test_loop(void)
487	{
488	long ecx, zf;
489	const long ecx_vals[] = {
490	0,
491	1,
492	0x10000,
493	0x10001,
494	#if defined(__x86_64__)
495	0x100000000L,
496	0x100000001L,
497	#endif
498	};
499	int i, res;
500
501	#if !defined(__x86_64__)
502	TEST_LOOP("jcxz");
503	TEST_LOOP("loopw");
504	TEST_LOOP("loopzw");
505	TEST_LOOP("loopnzw");
506	#endif
507
508	TEST_LOOP("jecxz");
509	TEST_LOOP("loopl");
510	TEST_LOOP("loopzl");
511	TEST_LOOP("loopnzl");
512	}
513
514	#undef CC_MASK
515	#ifdef TEST_P4_FLAGS
516	#define CC_MASK (CC_C \| CC_P \| CC_Z \| CC_S \| CC_O \| CC_A)
517	#else
518	#define CC_MASK (CC_O \| CC_C)
519	#endif
520
521	#define OP mul
522	#include "test-i386-muldiv.h"
523
524	#define OP imul
525	#include "test-i386-muldiv.h"
526
527	void test_imulw2(long op0, long op1)
528	{
529	long res, s1, s0, flags;
530	s0 = op0;
531	s1 = op1;
532	res = s0;
533	flags = 0;
534	asm volatile ("push %4\n\t"
535	"popf\n\t"
536	"imulw %w2, %w0\n\t"
537	"pushf\n\t"
538	"pop %1\n\t"
539	: "=q" (res), "=g" (flags)
540	: "q" (s1), "0" (res), "1" (flags));
541	printf("%-10s A=" FMTLX " B=" FMTLX " R=" FMTLX " CC=%04lx\n",
542	"imulw", s0, s1, res, flags & CC_MASK);
543	}
544
545	void test_imull2(long op0, long op1)
546	{
547	long res, s1, s0, flags;
548	s0 = op0;
549	s1 = op1;
550	res = s0;
551	flags = 0;
552	asm volatile ("push %4\n\t"
553	"popf\n\t"
554	"imull %k2, %k0\n\t"
555	"pushf\n\t"
556	"pop %1\n\t"
557	: "=q" (res), "=g" (flags)
558	: "q" (s1), "0" (res), "1" (flags));
559	printf("%-10s A=" FMTLX " B=" FMTLX " R=" FMTLX " CC=%04lx\n",
560	"imull", s0, s1, res, flags & CC_MASK);
561	}
562
563	#if defined(__x86_64__)
564	void test_imulq2(long op0, long op1)
565	{
566	long res, s1, s0, flags;
567	s0 = op0;
568	s1 = op1;
569	res = s0;
570	flags = 0;
571	asm volatile ("push %4\n\t"
572	"popf\n\t"
573	"imulq %2, %0\n\t"
574	"pushf\n\t"
575	"pop %1\n\t"
576	: "=q" (res), "=g" (flags)
577	: "q" (s1), "0" (res), "1" (flags));
578	printf("%-10s A=" FMTLX " B=" FMTLX " R=" FMTLX " CC=%04lx\n",
579	"imulq", s0, s1, res, flags & CC_MASK);
580	}
581	#endif
582
583	#define TEST_IMUL_IM(size, rsize, op0, op1)\
584	{\
585	long res, flags, s1;\
586	flags = 0;\
587	res = 0;\
588	s1 = op1;\
589	asm volatile ("push %3\n\t"\
590	"popf\n\t"\
591	"imul" size " $" #op0 ", %" rsize "2, %" rsize "0\n\t" \
592	"pushf\n\t"\
593	"pop %1\n\t"\
594	: "=r" (res), "=g" (flags)\
595	: "r" (s1), "1" (flags), "0" (res));\
596	printf("%-10s A=" FMTLX " B=" FMTLX " R=" FMTLX " CC=%04lx\n",\
597	"imul" size " im", (long)op0, (long)op1, res, flags & CC_MASK);\
598	}
599
600
601	#undef CC_MASK
602	#define CC_MASK (0)
603
604	#define OP div
605	#include "test-i386-muldiv.h"
606
607	#define OP idiv
608	#include "test-i386-muldiv.h"
609
610	void test_mul(void)
611	{
612	test_imulb(0x1234561d, 4);
613	test_imulb(3, -4);
614	test_imulb(0x80, 0x80);
615	test_imulb(0x10, 0x10);
616
617	test_imulw(0, 0x1234001d, 45);
618	test_imulw(0, 23, -45);
619	test_imulw(0, 0x8000, 0x8000);
620	test_imulw(0, 0x100, 0x100);
621
622	test_imull(0, 0x1234001d, 45);
623	test_imull(0, 23, -45);
624	test_imull(0, 0x80000000, 0x80000000);
625	test_imull(0, 0x10000, 0x10000);
626
627	test_mulb(0x1234561d, 4);
628	test_mulb(3, -4);
629	test_mulb(0x80, 0x80);
630	test_mulb(0x10, 0x10);
631
632	test_mulw(0, 0x1234001d, 45);
633	test_mulw(0, 23, -45);
634	test_mulw(0, 0x8000, 0x8000);
635	test_mulw(0, 0x100, 0x100);
636
637	test_mull(0, 0x1234001d, 45);
638	test_mull(0, 23, -45);
639	test_mull(0, 0x80000000, 0x80000000);
640	test_mull(0, 0x10000, 0x10000);
641
642	test_imulw2(0x1234001d, 45);
643	test_imulw2(23, -45);
644	test_imulw2(0x8000, 0x8000);
645	test_imulw2(0x100, 0x100);
646
647	test_imull2(0x1234001d, 45);
648	test_imull2(23, -45);
649	test_imull2(0x80000000, 0x80000000);
650	test_imull2(0x10000, 0x10000);
651
652	TEST_IMUL_IM("w", "w", 45, 0x1234);
653	TEST_IMUL_IM("w", "w", -45, 23);
654	TEST_IMUL_IM("w", "w", 0x8000, 0x80000000);
655	TEST_IMUL_IM("w", "w", 0x7fff, 0x1000);
656
657	TEST_IMUL_IM("l", "k", 45, 0x1234);
658	TEST_IMUL_IM("l", "k", -45, 23);
659	TEST_IMUL_IM("l", "k", 0x8000, 0x80000000);
660	TEST_IMUL_IM("l", "k", 0x7fff, 0x1000);
661
662	test_idivb(0x12341678, 0x127e);
663	test_idivb(0x43210123, -5);
664	test_idivb(0x12340004, -1);
665
666	test_idivw(0, 0x12345678, 12347);
667	test_idivw(0, -23223, -45);
668	test_idivw(0, 0x12348000, -1);
669	test_idivw(0x12343, 0x12345678, 0x81238567);
670
671	test_idivl(0, 0x12345678, 12347);
672	test_idivl(0, -233223, -45);
673	test_idivl(0, 0x80000000, -1);
674	test_idivl(0x12343, 0x12345678, 0x81234567);
675
676	test_divb(0x12341678, 0x127e);
677	test_divb(0x43210123, -5);
678	test_divb(0x12340004, -1);
679
680	test_divw(0, 0x12345678, 12347);
681	test_divw(0, -23223, -45);
682	test_divw(0, 0x12348000, -1);
683	test_divw(0x12343, 0x12345678, 0x81238567);
684
685	test_divl(0, 0x12345678, 12347);
686	test_divl(0, -233223, -45);
687	test_divl(0, 0x80000000, -1);
688	test_divl(0x12343, 0x12345678, 0x81234567);
689
690	#if defined(__x86_64__)
691	test_imulq(0, 0x1234001d1234001d, 45);
692	test_imulq(0, 23, -45);
693	test_imulq(0, 0x8000000000000000, 0x8000000000000000);
694	test_imulq(0, 0x100000000, 0x100000000);
695
696	test_mulq(0, 0x1234001d1234001d, 45);
697	test_mulq(0, 23, -45);
698	test_mulq(0, 0x8000000000000000, 0x8000000000000000);
699	test_mulq(0, 0x100000000, 0x100000000);
700
701	test_imulq2(0x1234001d1234001d, 45);
702	test_imulq2(23, -45);
703	test_imulq2(0x8000000000000000, 0x8000000000000000);
704	test_imulq2(0x100000000, 0x100000000);
705
706	TEST_IMUL_IM("q", "", 45, 0x12341234);
707	TEST_IMUL_IM("q", "", -45, 23);
708	TEST_IMUL_IM("q", "", 0x8000, 0x8000000000000000);
709	TEST_IMUL_IM("q", "", 0x7fff, 0x10000000);
710
711	test_idivq(0, 0x12345678abcdef, 12347);
712	test_idivq(0, -233223, -45);
713	test_idivq(0, 0x8000000000000000, -1);
714	test_idivq(0x12343, 0x12345678, 0x81234567);
715
716	test_divq(0, 0x12345678abcdef, 12347);
717	test_divq(0, -233223, -45);
718	test_divq(0, 0x8000000000000000, -1);
719	test_divq(0x12343, 0x12345678, 0x81234567);
720	#endif
721	}
722
723	#define TEST_BSX(op, size, op0)\
724	{\
725	long res, val, resz;\
726	val = op0;\
727	asm("xor %1, %1\n"\
728	"mov $0x12345678, %0\n"\
729	#op " %" size "2, %" size "0 ; setz %b1" \
730	: "=&r" (res), "=&q" (resz)\
731	: "r" (val));\
732	printf("%-10s A=" FMTLX " R=" FMTLX " %ld\n", #op, val, res, resz);\
733	}
734
735	void test_bsx(void)
736	{
737	TEST_BSX(bsrw, "w", 0);
738	TEST_BSX(bsrw, "w", 0x12340128);
739	TEST_BSX(bsfw, "w", 0);
740	TEST_BSX(bsfw, "w", 0x12340128);
741	TEST_BSX(bsrl, "k", 0);
742	TEST_BSX(bsrl, "k", 0x00340128);
743	TEST_BSX(bsfl, "k", 0);
744	TEST_BSX(bsfl, "k", 0x00340128);
745	#if defined(__x86_64__)
746	TEST_BSX(bsrq, "", 0);
747	TEST_BSX(bsrq, "", 0x003401281234);
748	TEST_BSX(bsfq, "", 0);
749	TEST_BSX(bsfq, "", 0x003401281234);
750	#endif
751	}
752
753	/**********************************************/
754
755	union float64u {
756	double d;
757	uint64_t l;
758	};
759
760	union float64u q_nan = { .l = 0xFFF8000000000000LL };
761	union float64u s_nan = { .l = 0xFFF0000000000000LL };
762
763	void test_fops(double a, double b)
764	{
765	printf("a=%f b=%f a+b=%f\n", a, b, a + b);
766	printf("a=%f b=%f a-b=%f\n", a, b, a - b);
767	printf("a=%f b=%f ab=%f\n", a, b, a b);
768	printf("a=%f b=%f a/b=%f\n", a, b, a / b);
769	printf("a=%f b=%f fmod(a, b)=%f\n", a, b, fmod(a, b));
770	printf("a=%f sqrt(a)=%f\n", a, sqrt(a));
771	printf("a=%f sin(a)=%f\n", a, sin(a));
772	printf("a=%f cos(a)=%f\n", a, cos(a));
773	printf("a=%f tan(a)=%f\n", a, tan(a));
774	printf("a=%f log(a)=%f\n", a, log(a));
775	printf("a=%f exp(a)=%f\n", a, exp(a));
776	printf("a=%f b=%f atan2(a, b)=%f\n", a, b, atan2(a, b));
777	/* just to test some op combining */
778	printf("a=%f asin(sin(a))=%f\n", a, asin(sin(a)));
779	printf("a=%f acos(cos(a))=%f\n", a, acos(cos(a)));
780	printf("a=%f atan(tan(a))=%f\n", a, atan(tan(a)));
781
782	}
783
784	void fpu_clear_exceptions(void)
785	{
786	struct __attribute__((packed)) {
787	uint16_t fpuc;
788	uint16_t dummy1;
789	uint16_t fpus;
790	uint16_t dummy2;
791	uint16_t fptag;
792	uint16_t dummy3;
793	uint32_t ignored[4];
794	long double fpregs[8];
795	} float_env32;
796
797	asm volatile ("fnstenv %0\n" : : "m" (float_env32));
798	float_env32.fpus &= ~0x7f;
799	asm volatile ("fldenv %0\n" : : "m" (float_env32));
800	}
801
802	/* XXX: display exception bits when supported */
803	#define FPUS_EMASK 0x0000
804	//#define FPUS_EMASK 0x007f
805
806	void test_fcmp(double a, double b)
807	{
808	long eflags, fpus;
809
810	fpu_clear_exceptions();
811	asm("fcom %2\n"
812	"fstsw %%ax\n"
813	: "=a" (fpus)
814	: "t" (a), "u" (b));
815	printf("fcom(%f %f)=%04lx \n",
816	a, b, fpus & (0x4500 \| FPUS_EMASK));
817	fpu_clear_exceptions();
818	asm("fucom %2\n"
819	"fstsw %%ax\n"
820	: "=a" (fpus)
821	: "t" (a), "u" (b));
822	printf("fucom(%f %f)=%04lx\n",
823	a, b, fpus & (0x4500 \| FPUS_EMASK));
824	if (TEST_FCOMI) {
825	/* test f(u)comi instruction */
826	fpu_clear_exceptions();
827	asm("fcomi %3, %2\n"
828	"fstsw %%ax\n"
829	"pushf\n"
830	"pop %0\n"
831	: "=r" (eflags), "=a" (fpus)
832	: "t" (a), "u" (b));
833	printf("fcomi(%f %f)=%04lx %02lx\n",
834	a, b, fpus & FPUS_EMASK, eflags & (CC_Z \| CC_P \| CC_C));
835	fpu_clear_exceptions();
836	asm("fucomi %3, %2\n"
837	"fstsw %%ax\n"
838	"pushf\n"
839	"pop %0\n"
840	: "=r" (eflags), "=a" (fpus)
841	: "t" (a), "u" (b));
842	printf("fucomi(%f %f)=%04lx %02lx\n",
843	a, b, fpus & FPUS_EMASK, eflags & (CC_Z \| CC_P \| CC_C));
844	}
845	fpu_clear_exceptions();
846	asm volatile("fxam\n"
847	"fstsw %%ax\n"
848	: "=a" (fpus)
849	: "t" (a));
850	printf("fxam(%f)=%04lx\n", a, fpus & 0x4700);
851	fpu_clear_exceptions();
852	}
853
854	void test_fcvt(double a)
855	{
856	float fa;
857	long double la;
858	int16_t fpuc;
859	int i;
860	int64_t lla;
861	int ia;
862	int16_t wa;
863	double ra;
864
865	fa = a;
866	la = a;
867	printf("(float)%f = %f\n", a, fa);
868	printf("(long double)%f = %Lf\n", a, la);
869	printf("a=" FMT64X "\n", (uint64_t )&a);
870	printf("la=" FMT64X " %04x\n", (uint64_t )&la,
871	(unsigned short )((char *)(&la) + 8));
872
873	/* test all roundings */
874	asm volatile ("fstcw %0" : "=m" (fpuc));
875	for(i=0;i<4;i++) {
876	uint16_t val16;
877	val16 = (fpuc & ~0x0c00) \| (i << 10);
878	asm volatile ("fldcw %0" : : "m" (val16));
879	asm volatile ("fist %0" : "=m" (wa) : "t" (a));
880	asm volatile ("fistl %0" : "=m" (ia) : "t" (a));
881	asm volatile ("fistpll %0" : "=m" (lla) : "t" (a) : "st");
882	asm volatile ("frndint ; fstl %0" : "=m" (ra) : "t" (a));
883	asm volatile ("fldcw %0" : : "m" (fpuc));
884	printf("(short)a = %d\n", wa);
885	printf("(int)a = %d\n", ia);
886	printf("(int64_t)a = " FMT64X "\n", lla);
887	printf("rint(a) = %f\n", ra);
888	}
889	}
890
891	#define TEST(N) \
892	asm("fld" #N : "=t" (a)); \
893	printf("fld" #N "= %f\n", a);
894
895	void test_fconst(void)
896	{
897	double a;
898	TEST(1);
899	TEST(l2t);
900	TEST(l2e);
901	TEST(pi);
902	TEST(lg2);
903	TEST(ln2);
904	TEST(z);
905	}
906
907	void test_fbcd(double a)
908	{
909	unsigned short bcd[5];
910	double b;
911
912	asm("fbstp %0" : "=m" (bcd[0]) : "t" (a) : "st");
913	asm("fbld %1" : "=t" (b) : "m" (bcd[0]));
914	printf("a=%f bcd=%04x%04x%04x%04x%04x b=%f\n",
915	a, bcd[4], bcd[3], bcd[2], bcd[1], bcd[0], b);
916	}
917
918	#define TEST_ENV(env, save, restore)\
919	{\
920	memset((env), 0xaa, sizeof(*(env)));\
921	for(i=0;i<5;i++)\
922	asm volatile ("fldl %0" : : "m" (dtab[i]));\
923	asm volatile (save " %0\n" : : "m" (*(env)));\
924	asm volatile (restore " %0\n": : "m" (*(env)));\
925	for(i=0;i<5;i++)\
926	asm volatile ("fstpl %0" : "=m" (rtab[i]));\
927	for(i=0;i<5;i++)\
928	printf("res[%d]=%f\n", i, rtab[i]);\
929	printf("fpuc=%04x fpus=%04x fptag=%04x\n",\
930	(env)->fpuc,\
931	(env)->fpus & 0xff00,\
932	(env)->fptag);\
933	}
934
935	void test_fenv(void)
936	{
937	struct __attribute__((packed)) {
938	uint16_t fpuc;
939	uint16_t dummy1;
940	uint16_t fpus;
941	uint16_t dummy2;
942	uint16_t fptag;
943	uint16_t dummy3;
944	uint32_t ignored[4];
945	long double fpregs[8];
946	} float_env32;
947	struct __attribute__((packed)) {
948	uint16_t fpuc;
949	uint16_t fpus;
950	uint16_t fptag;
951	uint16_t ignored[4];
952	long double fpregs[8];
953	} float_env16;
954	double dtab[8];
955	double rtab[8];
956	int i;
957
958	for(i=0;i<8;i++)
959	dtab[i] = i + 1;
960
961	TEST_ENV(&float_env16, "data16 fnstenv", "data16 fldenv");
962	TEST_ENV(&float_env16, "data16 fnsave", "data16 frstor");
963	TEST_ENV(&float_env32, "fnstenv", "fldenv");
964	TEST_ENV(&float_env32, "fnsave", "frstor");
965
966	/* test for ffree */
967	for(i=0;i<5;i++)
968	asm volatile ("fldl %0" : : "m" (dtab[i]));
969	asm volatile("ffree %st(2)");
970	asm volatile ("fnstenv %0\n" : : "m" (float_env32));
971	asm volatile ("fninit");
972	printf("fptag=%04x\n", float_env32.fptag);
973	}
974
975
976	#define TEST_FCMOV(a, b, eflags, CC)\
977	{\
978	double res;\
979	asm("push %3\n"\
980	"popf\n"\
981	"fcmov" CC " %2, %0\n"\
982	: "=t" (res)\
983	: "0" (a), "u" (b), "g" (eflags));\
984	printf("fcmov%s eflags=0x%04lx-> %f\n", \
985	CC, (long)eflags, res);\
986	}
987
988	void test_fcmov(void)
989	{
990	double a, b;
991	long eflags, i;
992
993	a = 1.0;
994	b = 2.0;
995	for(i = 0; i < 4; i++) {
996	eflags = 0;
997	if (i & 1)
998	eflags \|= CC_C;
999	if (i & 2)
1000	eflags \|= CC_Z;
1001	TEST_FCMOV(a, b, eflags, "b");
1002	TEST_FCMOV(a, b, eflags, "e");
1003	TEST_FCMOV(a, b, eflags, "be");
1004	TEST_FCMOV(a, b, eflags, "nb");
1005	TEST_FCMOV(a, b, eflags, "ne");
1006	TEST_FCMOV(a, b, eflags, "nbe");
1007	}
1008	TEST_FCMOV(a, b, 0, "u");
1009	TEST_FCMOV(a, b, CC_P, "u");
1010	TEST_FCMOV(a, b, 0, "nu");
1011	TEST_FCMOV(a, b, CC_P, "nu");
1012	}
1013
1014	void test_floats(void)
1015	{
1016	test_fops(2, 3);
1017	test_fops(1.4, -5);
1018	test_fcmp(2, -1);
1019	test_fcmp(2, 2);
1020	test_fcmp(2, 3);
1021	test_fcmp(2, q_nan.d);
1022	test_fcmp(q_nan.d, -1);
1023	test_fcmp(-1.0/0.0, -1);
1024	test_fcmp(1.0/0.0, -1);
1025	test_fcvt(0.5);
1026	test_fcvt(-0.5);
1027	test_fcvt(1.0/7.0);
1028	test_fcvt(-1.0/9.0);
1029	test_fcvt(32768);
1030	test_fcvt(-1e20);
1031	test_fcvt(-1.0/0.0);
1032	test_fcvt(1.0/0.0);
1033	test_fcvt(q_nan.d);
1034	test_fconst();
1035	test_fbcd(1234567890123456.0);
1036	test_fbcd(-123451234567890.0);
1037	test_fenv();
1038	if (TEST_CMOV) {
1039	test_fcmov();
1040	}
1041	}
1042
1043	/**********************************************/
1044	#if !defined(__x86_64__)
1045
1046	#define TEST_BCD(op, op0, cc_in, cc_mask)\
1047	{\
1048	int res, flags;\
1049	res = op0;\
1050	flags = cc_in;\
1051	asm ("push %3\n\t"\
1052	"popf\n\t"\
1053	#op "\n\t"\
1054	"pushf\n\t"\
1055	"pop %1\n\t"\
1056	: "=a" (res), "=g" (flags)\
1057	: "0" (res), "1" (flags));\
1058	printf("%-10s A=%08x R=%08x CCIN=%04x CC=%04x\n",\
1059	#op, op0, res, cc_in, flags & cc_mask);\
1060	}
1061
1062	void test_bcd(void)
1063	{
1064	TEST_BCD(daa, 0x12340503, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1065	TEST_BCD(daa, 0x12340506, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1066	TEST_BCD(daa, 0x12340507, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1067	TEST_BCD(daa, 0x12340559, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1068	TEST_BCD(daa, 0x12340560, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1069	TEST_BCD(daa, 0x1234059f, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1070	TEST_BCD(daa, 0x123405a0, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1071	TEST_BCD(daa, 0x12340503, 0, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1072	TEST_BCD(daa, 0x12340506, 0, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1073	TEST_BCD(daa, 0x12340503, CC_C, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1074	TEST_BCD(daa, 0x12340506, CC_C, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1075	TEST_BCD(daa, 0x12340503, CC_C \| CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1076	TEST_BCD(daa, 0x12340506, CC_C \| CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1077
1078	TEST_BCD(das, 0x12340503, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1079	TEST_BCD(das, 0x12340506, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1080	TEST_BCD(das, 0x12340507, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1081	TEST_BCD(das, 0x12340559, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1082	TEST_BCD(das, 0x12340560, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1083	TEST_BCD(das, 0x1234059f, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1084	TEST_BCD(das, 0x123405a0, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1085	TEST_BCD(das, 0x12340503, 0, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1086	TEST_BCD(das, 0x12340506, 0, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1087	TEST_BCD(das, 0x12340503, CC_C, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1088	TEST_BCD(das, 0x12340506, CC_C, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1089	TEST_BCD(das, 0x12340503, CC_C \| CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1090	TEST_BCD(das, 0x12340506, CC_C \| CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_A));
1091
1092	TEST_BCD(aaa, 0x12340205, CC_A, (CC_C \| CC_A));
1093	TEST_BCD(aaa, 0x12340306, CC_A, (CC_C \| CC_A));
1094	TEST_BCD(aaa, 0x1234040a, CC_A, (CC_C \| CC_A));
1095	TEST_BCD(aaa, 0x123405fa, CC_A, (CC_C \| CC_A));
1096	TEST_BCD(aaa, 0x12340205, 0, (CC_C \| CC_A));
1097	TEST_BCD(aaa, 0x12340306, 0, (CC_C \| CC_A));
1098	TEST_BCD(aaa, 0x1234040a, 0, (CC_C \| CC_A));
1099	TEST_BCD(aaa, 0x123405fa, 0, (CC_C \| CC_A));
1100
1101	TEST_BCD(aas, 0x12340205, CC_A, (CC_C \| CC_A));
1102	TEST_BCD(aas, 0x12340306, CC_A, (CC_C \| CC_A));
1103	TEST_BCD(aas, 0x1234040a, CC_A, (CC_C \| CC_A));
1104	TEST_BCD(aas, 0x123405fa, CC_A, (CC_C \| CC_A));
1105	TEST_BCD(aas, 0x12340205, 0, (CC_C \| CC_A));
1106	TEST_BCD(aas, 0x12340306, 0, (CC_C \| CC_A));
1107	TEST_BCD(aas, 0x1234040a, 0, (CC_C \| CC_A));
1108	TEST_BCD(aas, 0x123405fa, 0, (CC_C \| CC_A));
1109
1110	TEST_BCD(aam, 0x12340547, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_O \| CC_A));
1111	TEST_BCD(aad, 0x12340407, CC_A, (CC_C \| CC_P \| CC_Z \| CC_S \| CC_O \| CC_A));
1112	}
1113	#endif
1114
1115	#define TEST_XCHG(op, size, opconst)\
1116	{\
1117	long op0, op1;\
1118	op0 = i2l(0x12345678);\
1119	op1 = i2l(0xfbca7654);\
1120	asm(#op " %" size "0, %" size "1" \
1121	: "=q" (op0), opconst (op1) \
1122	: "0" (op0));\
1123	printf("%-10s A=" FMTLX " B=" FMTLX "\n",\
1124	#op, op0, op1);\
1125	}
1126
1127	#define TEST_CMPXCHG(op, size, opconst, eax)\
1128	{\
1129	long op0, op1, op2;\
1130	op0 = i2l(0x12345678);\
1131	op1 = i2l(0xfbca7654);\
1132	op2 = i2l(eax);\
1133	asm(#op " %" size "0, %" size "1" \
1134	: "=q" (op0), opconst (op1) \
1135	: "0" (op0), "a" (op2));\
1136	printf("%-10s EAX=" FMTLX " A=" FMTLX " C=" FMTLX "\n",\
1137	#op, op2, op0, op1);\
1138	}
1139
1140	void test_xchg(void)
1141	{
1142	#if defined(__x86_64__)
1143	TEST_XCHG(xchgq, "", "+q");
1144	#endif
1145	TEST_XCHG(xchgl, "k", "+q");
1146	TEST_XCHG(xchgw, "w", "+q");
1147	TEST_XCHG(xchgb, "b", "+q");
1148
1149	#if defined(__x86_64__)
1150	TEST_XCHG(xchgq, "", "=m");
1151	#endif
1152	TEST_XCHG(xchgl, "k", "+m");
1153	TEST_XCHG(xchgw, "w", "+m");
1154	TEST_XCHG(xchgb, "b", "+m");
1155
1156	#if defined(__x86_64__)
1157	TEST_XCHG(xaddq, "", "+q");
1158	#endif
1159	TEST_XCHG(xaddl, "k", "+q");
1160	TEST_XCHG(xaddw, "w", "+q");
1161	TEST_XCHG(xaddb, "b", "+q");
1162
1163	{
1164	int res;
1165	res = 0x12345678;
1166	asm("xaddl %1, %0" : "=r" (res) : "0" (res));
1167	printf("xaddl same res=%08x\n", res);
1168	}
1169
1170	#if defined(__x86_64__)
1171	TEST_XCHG(xaddq, "", "+m");
1172	#endif
1173	TEST_XCHG(xaddl, "k", "+m");
1174	TEST_XCHG(xaddw, "w", "+m");
1175	TEST_XCHG(xaddb, "b", "+m");
1176
1177	#if defined(__x86_64__)
1178	TEST_CMPXCHG(cmpxchgq, "", "+q", 0xfbca7654);
1179	#endif
1180	TEST_CMPXCHG(cmpxchgl, "k", "+q", 0xfbca7654);
1181	TEST_CMPXCHG(cmpxchgw, "w", "+q", 0xfbca7654);
1182	TEST_CMPXCHG(cmpxchgb, "b", "+q", 0xfbca7654);
1183
1184	#if defined(__x86_64__)
1185	TEST_CMPXCHG(cmpxchgq, "", "+q", 0xfffefdfc);
1186	#endif
1187	TEST_CMPXCHG(cmpxchgl, "k", "+q", 0xfffefdfc);
1188	TEST_CMPXCHG(cmpxchgw, "w", "+q", 0xfffefdfc);
1189	TEST_CMPXCHG(cmpxchgb, "b", "+q", 0xfffefdfc);
1190
1191	#if defined(__x86_64__)
1192	TEST_CMPXCHG(cmpxchgq, "", "+m", 0xfbca7654);
1193	#endif
1194	TEST_CMPXCHG(cmpxchgl, "k", "+m", 0xfbca7654);
1195	TEST_CMPXCHG(cmpxchgw, "w", "+m", 0xfbca7654);
1196	TEST_CMPXCHG(cmpxchgb, "b", "+m", 0xfbca7654);
1197
1198	#if defined(__x86_64__)
1199	TEST_CMPXCHG(cmpxchgq, "", "+m", 0xfffefdfc);
1200	#endif
1201	TEST_CMPXCHG(cmpxchgl, "k", "+m", 0xfffefdfc);
1202	TEST_CMPXCHG(cmpxchgw, "w", "+m", 0xfffefdfc);
1203	TEST_CMPXCHG(cmpxchgb, "b", "+m", 0xfffefdfc);
1204
1205	{
1206	uint64_t op0, op1, op2;
1207	long eax, edx;
1208	long i, eflags;
1209
1210	for(i = 0; i < 2; i++) {
1211	op0 = 0x123456789abcdLL;
1212	eax = i2l(op0 & 0xffffffff);
1213	edx = i2l(op0 >> 32);
1214	if (i == 0)
1215	op1 = 0xfbca765423456LL;
1216	else
1217	op1 = op0;
1218	op2 = 0x6532432432434LL;
1219	asm("cmpxchg8b %2\n"
1220	"pushf\n"
1221	"pop %3\n"
1222	: "=a" (eax), "=d" (edx), "=m" (op1), "=g" (eflags)
1223	: "0" (eax), "1" (edx), "m" (op1), "b" ((int)op2), "c" ((int)(op2 >> 32)));
1224	printf("cmpxchg8b: eax=" FMTLX " edx=" FMTLX " op1=" FMT64X " CC=%02lx\n",
1225	eax, edx, op1, eflags & CC_Z);
1226	}
1227	}
1228	}
1229
1230	#ifdef TEST_SEGS
1231	/**********************************************/
1232	/* segmentation tests */
1233
1234	#include <sys/syscall.h>
1235	#include <unistd.h>
1236	#include <asm/ldt.h>
1237	#include <linux/version.h>
1238
1239	static inline int modify_ldt(int func, void * ptr, unsigned long bytecount)
1240	{
1241	return syscall(__NR_modify_ldt, func, ptr, bytecount);
1242	}
1243
1244	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 5, 66)
1245	#define modify_ldt_ldt_s user_desc
1246	#endif
1247
1248	#define MK_SEL(n) (((n) << 3) \| 7)
1249
1250	uint8_t seg_data1[4096];
1251	uint8_t seg_data2[4096];
1252
1253	#define TEST_LR(op, size, seg, mask)\
1254	{\
1255	int res, res2;\
1256	uint16_t mseg = seg;\
1257	res = 0x12345678;\
1258	asm (op " %" size "2, %" size "0\n" \
1259	"movl $0, %1\n"\
1260	"jnz 1f\n"\
1261	"movl $1, %1\n"\
1262	"1:\n"\
1263	: "=r" (res), "=r" (res2) : "m" (mseg), "0" (res));\
1264	printf(op ": Z=%d %08x\n", res2, res & ~(mask));\
1265	}
1266
1267	#define TEST_ARPL(op, size, op1, op2)\
1268	{\
1269	long a, b, c; \
1270	a = (op1); \
1271	b = (op2); \
1272	asm volatile(op " %" size "3, %" size "0\n"\
1273	"movl $0,%1\n"\
1274	"jnz 1f\n"\
1275	"movl $1,%1\n"\
1276	"1:\n"\
1277	: "=r" (a), "=r" (c) : "0" (a), "r" (b)); \
1278	printf(op size " A=" FMTLX " B=" FMTLX " R=" FMTLX " z=%ld\n",\
1279	(long)(op1), (long)(op2), a, c);\
1280	}
1281
1282	/* NOTE: we use Linux modify_ldt syscall */
1283	void test_segs(void)
1284	{
1285	struct modify_ldt_ldt_s ldt;
1286	long long ldt_table[3];
1287	int res, res2;
1288	char tmp;
1289	struct {
1290	uint32_t offset;
1291	uint16_t seg;
1292	} __attribute__((packed)) segoff;
1293
1294	ldt.entry_number = 1;
1295	ldt.base_addr = (unsigned long)&seg_data1;
1296	ldt.limit = (sizeof(seg_data1) + 0xfff) >> 12;
1297	ldt.seg_32bit = 1;
1298	ldt.contents = MODIFY_LDT_CONTENTS_DATA;
1299	ldt.read_exec_only = 0;
1300	ldt.limit_in_pages = 1;
1301	ldt.seg_not_present = 0;
1302	ldt.useable = 1;
1303	modify_ldt(1, &ldt, sizeof(ldt)); /* write ldt entry */
1304
1305	ldt.entry_number = 2;
1306	ldt.base_addr = (unsigned long)&seg_data2;
1307	ldt.limit = (sizeof(seg_data2) + 0xfff) >> 12;
1308	ldt.seg_32bit = 1;
1309	ldt.contents = MODIFY_LDT_CONTENTS_DATA;
1310	ldt.read_exec_only = 0;
1311	ldt.limit_in_pages = 1;
1312	ldt.seg_not_present = 0;
1313	ldt.useable = 1;
1314	modify_ldt(1, &ldt, sizeof(ldt)); /* write ldt entry */
1315
1316	modify_ldt(0, &ldt_table, sizeof(ldt_table)); /* read ldt entries */
1317	#if 0
1318	{
1319	int i;
1320	for(i=0;i<3;i++)
1321	printf("%d: %016Lx\n", i, ldt_table[i]);
1322	}
1323	#endif
1324	/* do some tests with fs or gs */
1325	asm volatile ("movl %0, %%fs" : : "r" (MK_SEL(1)));
1326
1327	seg_data1[1] = 0xaa;
1328	seg_data2[1] = 0x55;
1329
1330	asm volatile ("fs movzbl 0x1, %0" : "=r" (res));
1331	printf("FS[1] = %02x\n", res);
1332
1333	asm volatile ("pushl %%gs\n"
1334	"movl %1, %%gs\n"
1335	"gs movzbl 0x1, %0\n"
1336	"popl %%gs\n"
1337	: "=r" (res)
1338	: "r" (MK_SEL(2)));
1339	printf("GS[1] = %02x\n", res);
1340
1341	/* tests with ds/ss (implicit segment case) */
1342	tmp = 0xa5;
1343	asm volatile ("pushl %%ebp\n\t"
1344	"pushl %%ds\n\t"
1345	"movl %2, %%ds\n\t"
1346	"movl %3, %%ebp\n\t"
1347	"movzbl 0x1, %0\n\t"
1348	"movzbl (%%ebp), %1\n\t"
1349	"popl %%ds\n\t"
1350	"popl %%ebp\n\t"
1351	: "=r" (res), "=r" (res2)
1352	: "r" (MK_SEL(1)), "r" (&tmp));
1353	printf("DS[1] = %02x\n", res);
1354	printf("SS[tmp] = %02x\n", res2);
1355
1356	segoff.seg = MK_SEL(2);
1357	segoff.offset = 0xabcdef12;
1358	asm volatile("lfs %2, %0\n\t"
1359	"movl %%fs, %1\n\t"
1360	: "=r" (res), "=g" (res2)
1361	: "m" (segoff));
1362	printf("FS:reg = %04x:%08x\n", res2, res);
1363
1364	TEST_LR("larw", "w", MK_SEL(2), 0x0100);
1365	TEST_LR("larl", "", MK_SEL(2), 0x0100);
1366	TEST_LR("lslw", "w", MK_SEL(2), 0);
1367	TEST_LR("lsll", "", MK_SEL(2), 0);
1368
1369	TEST_LR("larw", "w", 0xfff8, 0);
1370	TEST_LR("larl", "", 0xfff8, 0);
1371	TEST_LR("lslw", "w", 0xfff8, 0);
1372	TEST_LR("lsll", "", 0xfff8, 0);
1373
1374	TEST_ARPL("arpl", "w", 0x12345678 \| 3, 0x762123c \| 1);
1375	TEST_ARPL("arpl", "w", 0x12345678 \| 1, 0x762123c \| 3);
1376	TEST_ARPL("arpl", "w", 0x12345678 \| 1, 0x762123c \| 1);
1377	}
1378
1379	/* 16 bit code test */
1380	extern char code16_start, code16_end;
1381	extern char code16_func1;
1382	extern char code16_func2;
1383	extern char code16_func3;
1384
1385	void test_code16(void)
1386	{
1387	struct modify_ldt_ldt_s ldt;
1388	int res, res2;
1389
1390	/* build a code segment */
1391	ldt.entry_number = 1;
1392	ldt.base_addr = (unsigned long)&code16_start;
1393	ldt.limit = &code16_end - &code16_start;
1394	ldt.seg_32bit = 0;
1395	ldt.contents = MODIFY_LDT_CONTENTS_CODE;
1396	ldt.read_exec_only = 0;
1397	ldt.limit_in_pages = 0;
1398	ldt.seg_not_present = 0;
1399	ldt.useable = 1;
1400	modify_ldt(1, &ldt, sizeof(ldt)); /* write ldt entry */
1401
1402	/* call the first function */
1403	asm volatile ("lcall %1, %2"
1404	: "=a" (res)
1405	: "i" (MK_SEL(1)), "i" (&code16_func1): "memory", "cc");
1406	printf("func1() = 0x%08x\n", res);
1407	asm volatile ("lcall %2, %3"
1408	: "=a" (res), "=c" (res2)
1409	: "i" (MK_SEL(1)), "i" (&code16_func2): "memory", "cc");
1410	printf("func2() = 0x%08x spdec=%d\n", res, res2);
1411	asm volatile ("lcall %1, %2"
1412	: "=a" (res)
1413	: "i" (MK_SEL(1)), "i" (&code16_func3): "memory", "cc");
1414	printf("func3() = 0x%08x\n", res);
1415	}
1416	#endif
1417
1418	#if defined(__x86_64__)
1419	asm(".globl func_lret\n"
1420	"func_lret:\n"
1421	"movl $0x87654641, %eax\n"
1422	"lretq\n");
1423	#else
1424	asm(".globl func_lret\n"
1425	"func_lret:\n"
1426	"movl $0x87654321, %eax\n"
1427	"lret\n"
1428
1429	".globl func_iret\n"
1430	"func_iret:\n"
1431	"movl $0xabcd4321, %eax\n"
1432	"iret\n");
1433	#endif
1434
1435	extern char func_lret;
1436	extern char func_iret;
1437
1438	void test_misc(void)
1439	{
1440	char table[256];
1441	long res, i;
1442
1443	for(i=0;i<256;i++) table[i] = 256 - i;
1444	res = 0x12345678;
1445	asm ("xlat" : "=a" (res) : "b" (table), "0" (res));
1446	printf("xlat: EAX=" FMTLX "\n", res);
1447
1448	#if defined(__x86_64__)
1449	#if 0
1450	{
1451	/* XXX: see if Intel Core2 and AMD64 behavior really
1452	differ. Here we implemented the Intel way which is not
1453	compatible yet with QEMU. */
1454	static struct __attribute__((packed)) {
1455	uint64_t offset;
1456	uint16_t seg;
1457	} desc;
1458	long cs_sel;
1459
1460	asm volatile ("mov %%cs, %0" : "=r" (cs_sel));
1461
1462	asm volatile ("push %1\n"
1463	"call func_lret\n"
1464	: "=a" (res)
1465	: "r" (cs_sel) : "memory", "cc");
1466	printf("func_lret=" FMTLX "\n", res);
1467
1468	desc.offset = (long)&func_lret;
1469	desc.seg = cs_sel;
1470
1471	asm volatile ("xor %%rax, %%rax\n"
1472	"rex64 lcall *(%%rcx)\n"
1473	: "=a" (res)
1474	: "c" (&desc)
1475	: "memory", "cc");
1476	printf("func_lret2=" FMTLX "\n", res);
1477
1478	asm volatile ("push %2\n"
1479	"mov $ 1f, %%rax\n"
1480	"push %%rax\n"
1481	"rex64 ljmp *(%%rcx)\n"
1482	"1:\n"
1483	: "=a" (res)
1484	: "c" (&desc), "b" (cs_sel)
1485	: "memory", "cc");
1486	printf("func_lret3=" FMTLX "\n", res);
1487	}
1488	#endif
1489	#else
1490	asm volatile ("push %%cs ; call %1"
1491	: "=a" (res)
1492	: "m" (func_lret): "memory", "cc");
1493	printf("func_lret=" FMTLX "\n", res);
1494
1495	asm volatile ("pushf ; push %%cs ; call %1"
1496	: "=a" (res)
1497	: "m" (func_iret): "memory", "cc");
1498	printf("func_iret=" FMTLX "\n", res);
1499	#endif
1500
1501	#if defined(__x86_64__)
1502	/* specific popl test */
1503	asm volatile ("push $12345432 ; push $0x9abcdef ; pop (%%rsp) ; pop %0"
1504	: "=g" (res));
1505	printf("popl esp=" FMTLX "\n", res);
1506	#else
1507	/* specific popl test */
1508	asm volatile ("pushl $12345432 ; pushl $0x9abcdef ; popl (%%esp) ; popl %0"
1509	: "=g" (res));
1510	printf("popl esp=" FMTLX "\n", res);
1511
1512	/* specific popw test */
1513	asm volatile ("pushl $12345432 ; pushl $0x9abcdef ; popw (%%esp) ; addl $2, %%esp ; popl %0"
1514	: "=g" (res));
1515	printf("popw esp=" FMTLX "\n", res);
1516	#endif
1517	}
1518
1519	uint8_t str_buffer[4096];
1520
1521	#define TEST_STRING1(OP, size, DF, REP)\
1522	{\
1523	long esi, edi, eax, ecx, eflags;\
1524	\
1525	esi = (long)(str_buffer + sizeof(str_buffer) / 2);\
1526	edi = (long)(str_buffer + sizeof(str_buffer) / 2) + 16;\
1527	eax = i2l(0x12345678);\
1528	ecx = 17;\
1529	\
1530	asm volatile ("push $0\n\t"\
1531	"popf\n\t"\
1532	DF "\n\t"\
1533	REP #OP size "\n\t"\
1534	"cld\n\t"\
1535	"pushf\n\t"\
1536	"pop %4\n\t"\
1537	: "=S" (esi), "=D" (edi), "=a" (eax), "=c" (ecx), "=g" (eflags)\
1538	: "0" (esi), "1" (edi), "2" (eax), "3" (ecx));\
1539	printf("%-10s ESI=" FMTLX " EDI=" FMTLX " EAX=" FMTLX " ECX=" FMTLX " EFL=%04x\n",\
1540	REP #OP size, esi, edi, eax, ecx,\
1541	(int)(eflags & (CC_C \| CC_P \| CC_Z \| CC_S \| CC_O \| CC_A)));\
1542	}
1543
1544	#define TEST_STRING(OP, REP)\
1545	TEST_STRING1(OP, "b", "", REP);\
1546	TEST_STRING1(OP, "w", "", REP);\
1547	TEST_STRING1(OP, "l", "", REP);\
1548	X86_64_ONLY(TEST_STRING1(OP, "q", "", REP));\
1549	TEST_STRING1(OP, "b", "std", REP);\
1550	TEST_STRING1(OP, "w", "std", REP);\
1551	TEST_STRING1(OP, "l", "std", REP);\
1552	X86_64_ONLY(TEST_STRING1(OP, "q", "std", REP))
1553
1554	void test_string(void)
1555	{
1556	int i;
1557	for(i = 0;i < sizeof(str_buffer); i++)
1558	str_buffer[i] = i + 0x56;
1559	TEST_STRING(stos, "");
1560	TEST_STRING(stos, "rep ");
1561	TEST_STRING(lods, ""); /* to verify stos */
1562	TEST_STRING(lods, "rep ");
1563	TEST_STRING(movs, "");
1564	TEST_STRING(movs, "rep ");
1565	TEST_STRING(lods, ""); /* to verify stos */
1566
1567	/* XXX: better tests */
1568	TEST_STRING(scas, "");
1569	TEST_STRING(scas, "repz ");
1570	TEST_STRING(scas, "repnz ");
1571	TEST_STRING(cmps, "");
1572	TEST_STRING(cmps, "repz ");
1573	TEST_STRING(cmps, "repnz ");
1574	}
1575
1576	#ifdef TEST_VM86
1577	/* VM86 test */
1578
1579	static inline void set_bit(uint8_t *a, unsigned int bit)
1580	{
1581	a[bit / 8] \|= (1 << (bit % 8));
1582	}
1583
1584	static inline uint8_t *seg_to_linear(unsigned int seg, unsigned int reg)
1585	{
1586	return (uint8_t *)((seg << 4) + (reg & 0xffff));
1587	}
1588
1589	static inline void pushw(struct vm86_regs *r, int val)
1590	{
1591	r->esp = (r->esp & ~0xffff) \| ((r->esp - 2) & 0xffff);
1592	(uint16_t )seg_to_linear(r->ss, r->esp) = val;
1593	}
1594
1595	static inline int vm86(int func, struct vm86plus_struct *v86)
1596	{
1597	return syscall(__NR_vm86, func, v86);
1598	}
1599
1600	extern char vm86_code_start;
1601	extern char vm86_code_end;
1602
1603	#define VM86_CODE_CS 0x100
1604	#define VM86_CODE_IP 0x100
1605
1606	void test_vm86(void)
1607	{
1608	struct vm86plus_struct ctx;
1609	struct vm86_regs *r;
1610	uint8_t *vm86_mem;
1611	int seg, ret;
1612
1613	vm86_mem = mmap((void *)0x00000000, 0x110000,
1614	PROT_WRITE \| PROT_READ \| PROT_EXEC,
1615	MAP_FIXED \| MAP_ANON \| MAP_PRIVATE, -1, 0);
1616	if (vm86_mem == MAP_FAILED) {
1617	printf("ERROR: could not map vm86 memory");
1618	return;
1619	}
1620	memset(&ctx, 0, sizeof(ctx));
1621
1622	/* init basic registers */
1623	r = &ctx.regs;
1624	r->eip = VM86_CODE_IP;
1625	r->esp = 0xfffe;
1626	seg = VM86_CODE_CS;
1627	r->cs = seg;
1628	r->ss = seg;
1629	r->ds = seg;
1630	r->es = seg;
1631	r->fs = seg;
1632	r->gs = seg;
1633	r->eflags = VIF_MASK;
1634
1635	/* move code to proper address. We use the same layout as a .com
1636	dos program. */
1637	memcpy(vm86_mem + (VM86_CODE_CS << 4) + VM86_CODE_IP,
1638	&vm86_code_start, &vm86_code_end - &vm86_code_start);
1639
1640	/* mark int 0x21 as being emulated */
1641	set_bit((uint8_t *)&ctx.int_revectored, 0x21);
1642
1643	for(;;) {
1644	ret = vm86(VM86_ENTER, &ctx);
1645	switch(VM86_TYPE(ret)) {
1646	case VM86_INTx:
1647	{
1648	int int_num, ah, v;
1649
1650	int_num = VM86_ARG(ret);
1651	if (int_num != 0x21)
1652	goto unknown_int;
1653	ah = (r->eax >> 8) & 0xff;
1654	switch(ah) {
1655	case 0x00: /* exit */
1656	goto the_end;
1657	case 0x02: /* write char */
1658	{
1659	uint8_t c = r->edx;
1660	putchar(c);
1661	}
1662	break;
1663	case 0x09: /* write string */
1664	{
1665	uint8_t c, *ptr;
1666	ptr = seg_to_linear(r->ds, r->edx);
1667	for(;;) {
1668	c = *ptr++;
1669	if (c == '$')
1670	break;
1671	putchar(c);
1672	}
1673	r->eax = (r->eax & ~0xff) \| '$';
1674	}
1675	break;
1676	case 0xff: /* extension: write eflags number in edx */
1677	v = (int)r->edx;
1678	#ifndef LINUX_VM86_IOPL_FIX
1679	v &= ~0x3000;
1680	#endif
1681	printf("%08x\n", v);
1682	break;
1683	default:
1684	unknown_int:
1685	printf("unsupported int 0x%02x\n", int_num);
1686	goto the_end;
1687	}
1688	}
1689	break;
1690	case VM86_SIGNAL:
1691	/* a signal came, we just ignore that */
1692	break;
1693	case VM86_STI:
1694	break;
1695	default:
1696	printf("ERROR: unhandled vm86 return code (0x%x)\n", ret);
1697	goto the_end;
1698	}
1699	}
1700	the_end:
1701	printf("VM86 end\n");
1702	munmap(vm86_mem, 0x110000);
1703	}
1704	#endif
1705
1706	/* exception tests */
1707	#if defined(__i386__) && !defined(REG_EAX)
1708	#define REG_EAX EAX
1709	#define REG_EBX EBX
1710	#define REG_ECX ECX
1711	#define REG_EDX EDX
1712	#define REG_ESI ESI
1713	#define REG_EDI EDI
1714	#define REG_EBP EBP
1715	#define REG_ESP ESP
1716	#define REG_EIP EIP
1717	#define REG_EFL EFL
1718	#define REG_TRAPNO TRAPNO
1719	#define REG_ERR ERR
1720	#endif
1721
1722	#if defined(__x86_64__)
1723	#define REG_EIP REG_RIP
1724	#endif
1725
1726	jmp_buf jmp_env;
1727	int v1;
1728	int tab[2];
1729
1730	void sig_handler(int sig, siginfo_t info, void puc)
1731	{
1732	struct ucontext *uc = puc;
1733
1734	printf("si_signo=%d si_errno=%d si_code=%d",
1735	info->si_signo, info->si_errno, info->si_code);
1736	printf(" si_addr=0x%08lx",
1737	(unsigned long)info->si_addr);
1738	printf("\n");
1739
1740	printf("trapno=" FMTLX " err=" FMTLX,
1741	(long)uc->uc_mcontext.gregs[REG_TRAPNO],
1742	(long)uc->uc_mcontext.gregs[REG_ERR]);
1743	printf(" EIP=" FMTLX, (long)uc->uc_mcontext.gregs[REG_EIP]);
1744	printf("\n");
1745	longjmp(jmp_env, 1);
1746	}
1747
1748	void test_exceptions(void)
1749	{
1750	struct sigaction act;
1751	volatile int val;
1752
1753	act.sa_sigaction = sig_handler;
1754	sigemptyset(&act.sa_mask);
1755	act.sa_flags = SA_SIGINFO \| SA_NODEFER;
1756	sigaction(SIGFPE, &act, NULL);
1757	sigaction(SIGILL, &act, NULL);
1758	sigaction(SIGSEGV, &act, NULL);
1759	sigaction(SIGBUS, &act, NULL);
1760	sigaction(SIGTRAP, &act, NULL);
1761
1762	/* test division by zero reporting */
1763	printf("DIVZ exception:\n");
1764	if (setjmp(jmp_env) == 0) {
1765	/* now divide by zero */
1766	v1 = 0;
1767	v1 = 2 / v1;
1768	}
1769
1770	#if !defined(__x86_64__)
1771	printf("BOUND exception:\n");
1772	if (setjmp(jmp_env) == 0) {
1773	/* bound exception */
1774	tab[0] = 1;
1775	tab[1] = 10;
1776	asm volatile ("bound %0, %1" : : "r" (11), "m" (tab[0]));
1777	}
1778	#endif
1779
1780	#ifdef TEST_SEGS
1781	printf("segment exceptions:\n");
1782	if (setjmp(jmp_env) == 0) {
1783	/* load an invalid segment */
1784	asm volatile ("movl %0, %%fs" : : "r" ((0x1234 << 3) \| 1));
1785	}
1786	if (setjmp(jmp_env) == 0) {
1787	/* null data segment is valid */
1788	asm volatile ("movl %0, %%fs" : : "r" (3));
1789	/* null stack segment */
1790	asm volatile ("movl %0, %%ss" : : "r" (3));
1791	}
1792
1793	{
1794	struct modify_ldt_ldt_s ldt;
1795	ldt.entry_number = 1;
1796	ldt.base_addr = (unsigned long)&seg_data1;
1797	ldt.limit = (sizeof(seg_data1) + 0xfff) >> 12;
1798	ldt.seg_32bit = 1;
1799	ldt.contents = MODIFY_LDT_CONTENTS_DATA;
1800	ldt.read_exec_only = 0;
1801	ldt.limit_in_pages = 1;
1802	ldt.seg_not_present = 1;
1803	ldt.useable = 1;
1804	modify_ldt(1, &ldt, sizeof(ldt)); /* write ldt entry */
1805
1806	if (setjmp(jmp_env) == 0) {
1807	/* segment not present */
1808	asm volatile ("movl %0, %%fs" : : "r" (MK_SEL(1)));
1809	}
1810	}
1811	#endif
1812
1813	/* test SEGV reporting */
1814	printf("PF exception:\n");
1815	if (setjmp(jmp_env) == 0) {
1816	val = 1;
1817	/* we add a nop to test a weird PC retrieval case */
1818	asm volatile ("nop");
1819	/* now store in an invalid address */
1820	(char )0x1234 = 1;
1821	}
1822
1823	/* test SEGV reporting */
1824	printf("PF exception:\n");
1825	if (setjmp(jmp_env) == 0) {
1826	val = 1;
1827	/* read from an invalid address */
1828	v1 = (char )0x1234;
1829	}
1830
1831	/* test illegal instruction reporting */
1832	printf("UD2 exception:\n");
1833	if (setjmp(jmp_env) == 0) {
1834	/* now execute an invalid instruction */
1835	asm volatile("ud2");
1836	}
1837	printf("lock nop exception:\n");
1838	if (setjmp(jmp_env) == 0) {
1839	/* now execute an invalid instruction */
1840	asm volatile("lock nop");
1841	}
1842
1843	printf("INT exception:\n");
1844	if (setjmp(jmp_env) == 0) {
1845	asm volatile ("int $0xfd");
1846	}
1847	if (setjmp(jmp_env) == 0) {
1848	asm volatile ("int $0x01");
1849	}
1850	if (setjmp(jmp_env) == 0) {
1851	asm volatile (".byte 0xcd, 0x03");
1852	}
1853	if (setjmp(jmp_env) == 0) {
1854	asm volatile ("int $0x04");
1855	}
1856	if (setjmp(jmp_env) == 0) {
1857	asm volatile ("int $0x05");
1858	}
1859
1860	printf("INT3 exception:\n");
1861	if (setjmp(jmp_env) == 0) {
1862	asm volatile ("int3");
1863	}
1864
1865	printf("CLI exception:\n");
1866	if (setjmp(jmp_env) == 0) {
1867	asm volatile ("cli");
1868	}
1869
1870	printf("STI exception:\n");
1871	if (setjmp(jmp_env) == 0) {
1872	asm volatile ("cli");
1873	}
1874
1875	#if !defined(__x86_64__)
1876	printf("INTO exception:\n");
1877	if (setjmp(jmp_env) == 0) {
1878	/* overflow exception */
1879	asm volatile ("addl $1, %0 ; into" : : "r" (0x7fffffff));
1880	}
1881	#endif
1882
1883	printf("OUTB exception:\n");
1884	if (setjmp(jmp_env) == 0) {
1885	asm volatile ("outb %%al, %%dx" : : "d" (0x4321), "a" (0));
1886	}
1887
1888	printf("INB exception:\n");
1889	if (setjmp(jmp_env) == 0) {
1890	asm volatile ("inb %%dx, %%al" : "=a" (val) : "d" (0x4321));
1891	}
1892
1893	printf("REP OUTSB exception:\n");
1894	if (setjmp(jmp_env) == 0) {
1895	asm volatile ("rep outsb" : : "d" (0x4321), "S" (tab), "c" (1));
1896	}
1897
1898	printf("REP INSB exception:\n");
1899	if (setjmp(jmp_env) == 0) {
1900	asm volatile ("rep insb" : : "d" (0x4321), "D" (tab), "c" (1));
1901	}
1902
1903	printf("HLT exception:\n");
1904	if (setjmp(jmp_env) == 0) {
1905	asm volatile ("hlt");
1906	}
1907
1908	printf("single step exception:\n");
1909	val = 0;
1910	if (setjmp(jmp_env) == 0) {
1911	asm volatile ("pushf\n"
1912	"orl $0x00100, (%%esp)\n"
1913	"popf\n"
1914	"movl $0xabcd, %0\n"
1915	"movl $0x0, %0\n" : "=m" (val) : : "cc", "memory");
1916	}
1917	printf("val=0x%x\n", val);
1918	}
1919
1920	#if !defined(__x86_64__)
1921	/* specific precise single step test */
1922	void sig_trap_handler(int sig, siginfo_t info, void puc)
1923	{
1924	struct ucontext *uc = puc;
1925	printf("EIP=" FMTLX "\n", (long)uc->uc_mcontext.gregs[REG_EIP]);
1926	}
1927
1928	const uint8_t sstep_buf1[4] = { 1, 2, 3, 4};
1929	uint8_t sstep_buf2[4];
1930
1931	void test_single_step(void)
1932	{
1933	struct sigaction act;
1934	volatile int val;
1935	int i;
1936
1937	val = 0;
1938	act.sa_sigaction = sig_trap_handler;
1939	sigemptyset(&act.sa_mask);
1940	act.sa_flags = SA_SIGINFO;
1941	sigaction(SIGTRAP, &act, NULL);
1942	asm volatile ("pushf\n"
1943	"orl $0x00100, (%%esp)\n"
1944	"popf\n"
1945	"movl $0xabcd, %0\n"
1946
1947	/* jmp test */
1948	"movl $3, %%ecx\n"
1949	"1:\n"
1950	"addl $1, %0\n"
1951	"decl %%ecx\n"
1952	"jnz 1b\n"
1953
1954	/* movsb: the single step should stop at each movsb iteration */
1955	"movl $sstep_buf1, %%esi\n"
1956	"movl $sstep_buf2, %%edi\n"
1957	"movl $0, %%ecx\n"
1958	"rep movsb\n"
1959	"movl $3, %%ecx\n"
1960	"rep movsb\n"
1961	"movl $1, %%ecx\n"
1962	"rep movsb\n"
1963
1964	/* cmpsb: the single step should stop at each cmpsb iteration */
1965	"movl $sstep_buf1, %%esi\n"
1966	"movl $sstep_buf2, %%edi\n"
1967	"movl $0, %%ecx\n"
1968	"rep cmpsb\n"
1969	"movl $4, %%ecx\n"
1970	"rep cmpsb\n"
1971
1972	/* getpid() syscall: single step should skip one
1973	instruction */
1974	"movl $20, %%eax\n"
1975	"int $0x80\n"
1976	"movl $0, %%eax\n"
1977
1978	/* when modifying SS, trace is not done on the next
1979	instruction */
1980	"movl %%ss, %%ecx\n"
1981	"movl %%ecx, %%ss\n"
1982	"addl $1, %0\n"
1983	"movl $1, %%eax\n"
1984	"movl %%ecx, %%ss\n"
1985	"jmp 1f\n"
1986	"addl $1, %0\n"
1987	"1:\n"
1988	"movl $1, %%eax\n"
1989	"pushl %%ecx\n"
1990	"popl %%ss\n"
1991	"addl $1, %0\n"
1992	"movl $1, %%eax\n"
1993
1994	"pushf\n"
1995	"andl $~0x00100, (%%esp)\n"
1996	"popf\n"
1997	: "=m" (val)
1998	:
1999	: "cc", "memory", "eax", "ecx", "esi", "edi");
2000	printf("val=%d\n", val);
2001	for(i = 0; i < 4; i++)
2002	printf("sstep_buf2[%d] = %d\n", i, sstep_buf2[i]);
2003	}
2004
2005	/* self modifying code test */
2006	uint8_t code[] = {
2007	0xb8, 0x1, 0x00, 0x00, 0x00, /* movl $1, %eax */
2008	0xc3, /* ret */
2009	};
2010
2011	asm(".section \".data\"\n"
2012	"smc_code2:\n"
2013	"movl 4(%esp), %eax\n"
2014	"movl %eax, smc_patch_addr2 + 1\n"
2015	"nop\n"
2016	"nop\n"
2017	"nop\n"
2018	"nop\n"
2019	"nop\n"
2020	"nop\n"
2021	"nop\n"
2022	"nop\n"
2023	"smc_patch_addr2:\n"
2024	"movl $1, %eax\n"
2025	"ret\n"
2026	".previous\n"
2027	);
2028
2029	typedef int FuncType(void);
2030	extern int smc_code2(int);
2031	void test_self_modifying_code(void)
2032	{
2033	int i;
2034	printf("self modifying code:\n");
2035	printf("func1 = 0x%x\n", ((FuncType *)code)());
2036	for(i = 2; i <= 4; i++) {
2037	code[1] = i;
2038	printf("func%d = 0x%x\n", i, ((FuncType *)code)());
2039	}
2040
2041	/* more difficult test : the modified code is just after the
2042	modifying instruction. It is forbidden in Intel specs, but it
2043	is used by old DOS programs */
2044	for(i = 2; i <= 4; i++) {
2045	printf("smc_code2(%d) = %d\n", i, smc_code2(i));
2046	}
2047	}
2048	#endif
2049
2050	long enter_stack[4096];
2051
2052	#if defined(__x86_64__)
2053	#define RSP "%%rsp"
2054	#define RBP "%%rbp"
2055	#else
2056	#define RSP "%%esp"
2057	#define RBP "%%ebp"
2058	#endif
2059
2060	#define TEST_ENTER(size, stack_type, level)\
2061	{\
2062	long esp_save, esp_val, ebp_val, ebp_save, i;\
2063	stack_type ptr, stack_end, *stack_ptr;\
2064	memset(enter_stack, 0, sizeof(enter_stack));\
2065	stack_end = stack_ptr = (stack_type *)(enter_stack + 4096);\
2066	ebp_val = (long)stack_ptr;\
2067	for(i=1;i<=32;i++)\
2068	*--stack_ptr = i;\
2069	esp_val = (long)stack_ptr;\
2070	asm("mov " RSP ", %[esp_save]\n"\
2071	"mov " RBP ", %[ebp_save]\n"\
2072	"mov %[esp_val], " RSP "\n"\
2073	"mov %[ebp_val], " RBP "\n"\
2074	"enter" size " $8, $" #level "\n"\
2075	"mov " RSP ", %[esp_val]\n"\
2076	"mov " RBP ", %[ebp_val]\n"\
2077	"mov %[esp_save], " RSP "\n"\
2078	"mov %[ebp_save], " RBP "\n"\
2079	: [esp_save] "=r" (esp_save),\
2080	[ebp_save] "=r" (ebp_save),\
2081	[esp_val] "=r" (esp_val),\
2082	[ebp_val] "=r" (ebp_val)\
2083	: "[esp_val]" (esp_val),\
2084	"[ebp_val]" (ebp_val));\
2085	printf("level=%d:\n", level);\
2086	printf("esp_val=" FMTLX "\n", esp_val - (long)stack_end);\
2087	printf("ebp_val=" FMTLX "\n", ebp_val - (long)stack_end);\
2088	for(ptr = (stack_type *)esp_val; ptr < stack_end; ptr++)\
2089	printf(FMTLX "\n", (long)ptr[0]);\
2090	}
2091
2092	static void test_enter(void)
2093	{
2094	#if defined(__x86_64__)
2095	TEST_ENTER("q", uint64_t, 0);
2096	TEST_ENTER("q", uint64_t, 1);
2097	TEST_ENTER("q", uint64_t, 2);
2098	TEST_ENTER("q", uint64_t, 31);
2099	#else
2100	TEST_ENTER("l", uint32_t, 0);
2101	TEST_ENTER("l", uint32_t, 1);
2102	TEST_ENTER("l", uint32_t, 2);
2103	TEST_ENTER("l", uint32_t, 31);
2104	#endif
2105
2106	TEST_ENTER("w", uint16_t, 0);
2107	TEST_ENTER("w", uint16_t, 1);
2108	TEST_ENTER("w", uint16_t, 2);
2109	TEST_ENTER("w", uint16_t, 31);
2110	}
2111
2112	#ifdef TEST_SSE
2113
2114	typedef int __m64 __attribute__ ((__mode__ (__V2SI__)));
2115	typedef float __m128 __attribute__ ((__mode__(__V4SF__)));
2116
2117	typedef union {
2118	double d[2];
2119	float s[4];
2120	uint32_t l[4];
2121	uint64_t q[2];
2122	__m128 dq;
2123	} XMMReg;
2124
2125	static uint64_t __attribute__((aligned(16))) test_values[4][2] = {
2126	{ 0x456723c698694873, 0xdc515cff944a58ec },
2127	{ 0x1f297ccd58bad7ab, 0x41f21efba9e3e146 },
2128	{ 0x007c62c2085427f8, 0x231be9e8cde7438d },
2129	{ 0x0f76255a085427f8, 0xc233e9e8c4c9439a },
2130	};
2131
2132	#define SSE_OP(op)\
2133	{\
2134	asm volatile (#op " %2, %0" : "=x" (r.dq) : "0" (a.dq), "x" (b.dq));\
2135	printf("%-9s: a=" FMT64X "" FMT64X " b=" FMT64X "" FMT64X " r=" FMT64X "" FMT64X "\n",\
2136	#op,\
2137	a.q[1], a.q[0],\
2138	b.q[1], b.q[0],\
2139	r.q[1], r.q[0]);\
2140	}
2141
2142	#define SSE_OP2(op)\
2143	{\
2144	int i;\
2145	for(i=0;i<2;i++) {\
2146	a.q[0] = test_values[2*i][0];\
2147	a.q[1] = test_values[2*i][1];\
2148	b.q[0] = test_values[2*i+1][0];\
2149	b.q[1] = test_values[2*i+1][1];\
2150	SSE_OP(op);\
2151	}\
2152	}
2153
2154	#define MMX_OP2(op)\
2155	{\
2156	int i;\
2157	for(i=0;i<2;i++) {\
2158	a.q[0] = test_values[2*i][0];\
2159	b.q[0] = test_values[2*i+1][0];\
2160	asm volatile (#op " %2, %0" : "=y" (r.q[0]) : "0" (a.q[0]), "y" (b.q[0]));\
2161	printf("%-9s: a=" FMT64X " b=" FMT64X " r=" FMT64X "\n",\
2162	#op,\
2163	a.q[0],\
2164	b.q[0],\
2165	r.q[0]);\
2166	}\
2167	SSE_OP2(op);\
2168	}
2169
2170	#define SHUF_OP(op, ib)\
2171	{\
2172	a.q[0] = test_values[0][0];\
2173	a.q[1] = test_values[0][1];\
2174	b.q[0] = test_values[1][0];\
2175	b.q[1] = test_values[1][1];\
2176	asm volatile (#op " $" #ib ", %2, %0" : "=x" (r.dq) : "0" (a.dq), "x" (b.dq));\
2177	printf("%-9s: a=" FMT64X "" FMT64X " b=" FMT64X "" FMT64X " ib=%02x r=" FMT64X "" FMT64X "\n",\
2178	#op,\
2179	a.q[1], a.q[0],\
2180	b.q[1], b.q[0],\
2181	ib,\
2182	r.q[1], r.q[0]);\
2183	}
2184
2185	#define PSHUF_OP(op, ib)\
2186	{\
2187	int i;\
2188	for(i=0;i<2;i++) {\
2189	a.q[0] = test_values[2*i][0];\
2190	a.q[1] = test_values[2*i][1];\
2191	asm volatile (#op " $" #ib ", %1, %0" : "=x" (r.dq) : "x" (a.dq));\
2192	printf("%-9s: a=" FMT64X "" FMT64X " ib=%02x r=" FMT64X "" FMT64X "\n",\
2193	#op,\
2194	a.q[1], a.q[0],\
2195	ib,\
2196	r.q[1], r.q[0]);\
2197	}\
2198	}
2199
2200	#define SHIFT_IM(op, ib)\
2201	{\
2202	int i;\
2203	for(i=0;i<2;i++) {\
2204	a.q[0] = test_values[2*i][0];\
2205	a.q[1] = test_values[2*i][1];\
2206	asm volatile (#op " $" #ib ", %0" : "=x" (r.dq) : "0" (a.dq));\
2207	printf("%-9s: a=" FMT64X "" FMT64X " ib=%02x r=" FMT64X "" FMT64X "\n",\
2208	#op,\
2209	a.q[1], a.q[0],\
2210	ib,\
2211	r.q[1], r.q[0]);\
2212	}\
2213	}
2214
2215	#define SHIFT_OP(op, ib)\
2216	{\
2217	int i;\
2218	SHIFT_IM(op, ib);\
2219	for(i=0;i<2;i++) {\
2220	a.q[0] = test_values[2*i][0];\
2221	a.q[1] = test_values[2*i][1];\
2222	b.q[0] = ib;\
2223	b.q[1] = 0;\
2224	asm volatile (#op " %2, %0" : "=x" (r.dq) : "0" (a.dq), "x" (b.dq));\
2225	printf("%-9s: a=" FMT64X "" FMT64X " b=" FMT64X "" FMT64X " r=" FMT64X "" FMT64X "\n",\
2226	#op,\
2227	a.q[1], a.q[0],\
2228	b.q[1], b.q[0],\
2229	r.q[1], r.q[0]);\
2230	}\
2231	}
2232
2233	#define MOVMSK(op)\
2234	{\
2235	int i, reg;\
2236	for(i=0;i<2;i++) {\
2237	a.q[0] = test_values[2*i][0];\
2238	a.q[1] = test_values[2*i][1];\
2239	asm volatile (#op " %1, %0" : "=r" (reg) : "x" (a.dq));\
2240	printf("%-9s: a=" FMT64X "" FMT64X " r=%08x\n",\
2241	#op,\
2242	a.q[1], a.q[0],\
2243	reg);\
2244	}\
2245	}
2246
2247	#define SSE_OPS(a) \
2248	SSE_OP(a ## ps);\
2249	SSE_OP(a ## ss);
2250
2251	#define SSE_OPD(a) \
2252	SSE_OP(a ## pd);\
2253	SSE_OP(a ## sd);
2254
2255	#define SSE_COMI(op, field)\
2256	{\
2257	unsigned int eflags;\
2258	XMMReg a, b;\
2259	a.field[0] = a1;\
2260	b.field[0] = b1;\
2261	asm volatile (#op " %2, %1\n"\
2262	"pushf\n"\
2263	"pop %0\n"\
2264	: "=m" (eflags)\
2265	: "x" (a.dq), "x" (b.dq));\
2266	printf("%-9s: a=%f b=%f cc=%04x\n",\
2267	#op, a1, b1,\
2268	eflags & (CC_C \| CC_P \| CC_Z \| CC_S \| CC_O \| CC_A));\
2269	}
2270
2271	void test_sse_comi(double a1, double b1)
2272	{
2273	SSE_COMI(ucomiss, s);
2274	SSE_COMI(ucomisd, d);
2275	SSE_COMI(comiss, s);
2276	SSE_COMI(comisd, d);
2277	}
2278
2279	#define CVT_OP_XMM(op)\
2280	{\
2281	asm volatile (#op " %1, %0" : "=x" (r.dq) : "x" (a.dq));\
2282	printf("%-9s: a=" FMT64X "" FMT64X " r=" FMT64X "" FMT64X "\n",\
2283	#op,\
2284	a.q[1], a.q[0],\
2285	r.q[1], r.q[0]);\
2286	}
2287
2288	/* Force %xmm0 usage to avoid the case where both register index are 0
2289	to test intruction decoding more extensively */
2290	#define CVT_OP_XMM2MMX(op)\
2291	{\
2292	asm volatile (#op " %1, %0" : "=y" (r.q[0]) : "x" (a.dq) \
2293	: "%xmm0"); \
2294	asm volatile("emms\n"); \
2295	printf("%-9s: a=" FMT64X "" FMT64X " r=" FMT64X "\n",\
2296	#op,\
2297	a.q[1], a.q[0],\
2298	r.q[0]);\
2299	}
2300
2301	#define CVT_OP_MMX2XMM(op)\
2302	{\
2303	asm volatile (#op " %1, %0" : "=x" (r.dq) : "y" (a.q[0]));\
2304	asm volatile("emms\n"); \
2305	printf("%-9s: a=" FMT64X " r=" FMT64X "" FMT64X "\n",\
2306	#op,\
2307	a.q[0],\
2308	r.q[1], r.q[0]);\
2309	}
2310
2311	#define CVT_OP_REG2XMM(op)\
2312	{\
2313	asm volatile (#op " %1, %0" : "=x" (r.dq) : "r" (a.l[0]));\
2314	printf("%-9s: a=%08x r=" FMT64X "" FMT64X "\n",\
2315	#op,\
2316	a.l[0],\
2317	r.q[1], r.q[0]);\
2318	}
2319
2320	#define CVT_OP_XMM2REG(op)\
2321	{\
2322	asm volatile (#op " %1, %0" : "=r" (r.l[0]) : "x" (a.dq));\
2323	printf("%-9s: a=" FMT64X "" FMT64X " r=%08x\n",\
2324	#op,\
2325	a.q[1], a.q[0],\
2326	r.l[0]);\
2327	}
2328
2329	struct fpxstate {
2330	uint16_t fpuc;
2331	uint16_t fpus;
2332	uint16_t fptag;
2333	uint16_t fop;
2334	uint32_t fpuip;
2335	uint16_t cs_sel;
2336	uint16_t dummy0;
2337	uint32_t fpudp;
2338	uint16_t ds_sel;
2339	uint16_t dummy1;
2340	uint32_t mxcsr;
2341	uint32_t mxcsr_mask;
2342	uint8_t fpregs1[8 * 16];
2343	uint8_t xmm_regs[8 * 16];
2344	uint8_t dummy2[224];
2345	};
2346
2347	static struct fpxstate fpx_state __attribute__((aligned(16)));
2348	static struct fpxstate fpx_state2 __attribute__((aligned(16)));
2349
2350	void test_fxsave(void)
2351	{
2352	struct fpxstate *fp = &fpx_state;
2353	struct fpxstate *fp2 = &fpx_state2;
2354	int i, nb_xmm;
2355	XMMReg a, b;
2356	a.q[0] = test_values[0][0];
2357	a.q[1] = test_values[0][1];
2358	b.q[0] = test_values[1][0];
2359	b.q[1] = test_values[1][1];
2360
2361	asm("movdqa %2, %%xmm0\n"
2362	"movdqa %3, %%xmm7\n"
2363	#if defined(__x86_64__)
2364	"movdqa %2, %%xmm15\n"
2365	#endif
2366	" fld1\n"
2367	" fldpi\n"
2368	" fldln2\n"
2369	" fxsave %0\n"
2370	" fxrstor %0\n"
2371	" fxsave %1\n"
2372	" fninit\n"
2373	: "=m" ((uint32_t )fp2), "=m" ((uint32_t )fp)
2374	: "m" (a), "m" (b));
2375	printf("fpuc=%04x\n", fp->fpuc);
2376	printf("fpus=%04x\n", fp->fpus);
2377	printf("fptag=%04x\n", fp->fptag);
2378	for(i = 0; i < 3; i++) {
2379	printf("ST%d: " FMT64X " %04x\n",
2380	i,
2381	(uint64_t )&fp->fpregs1[i * 16],
2382	(uint16_t )&fp->fpregs1[i * 16 + 8]);
2383	}
2384	printf("mxcsr=%08x\n", fp->mxcsr & 0x1f80);
2385	#if defined(__x86_64__)
2386	nb_xmm = 16;
2387	#else
2388	nb_xmm = 8;
2389	#endif
2390	for(i = 0; i < nb_xmm; i++) {
2391	printf("xmm%d: " FMT64X "" FMT64X "\n",
2392	i,
2393	(uint64_t )&fp->xmm_regs[i * 16],
2394	(uint64_t )&fp->xmm_regs[i * 16 + 8]);
2395	}
2396	}
2397
2398	void test_sse(void)
2399	{
2400	XMMReg r, a, b;
2401	int i;
2402
2403	MMX_OP2(punpcklbw);
2404	MMX_OP2(punpcklwd);
2405	MMX_OP2(punpckldq);
2406	MMX_OP2(packsswb);
2407	MMX_OP2(pcmpgtb);
2408	MMX_OP2(pcmpgtw);
2409	MMX_OP2(pcmpgtd);
2410	MMX_OP2(packuswb);
2411	MMX_OP2(punpckhbw);
2412	MMX_OP2(punpckhwd);
2413	MMX_OP2(punpckhdq);
2414	MMX_OP2(packssdw);
2415	MMX_OP2(pcmpeqb);
2416	MMX_OP2(pcmpeqw);
2417	MMX_OP2(pcmpeqd);
2418
2419	MMX_OP2(paddq);
2420	MMX_OP2(pmullw);
2421	MMX_OP2(psubusb);
2422	MMX_OP2(psubusw);
2423	MMX_OP2(pminub);
2424	MMX_OP2(pand);
2425	MMX_OP2(paddusb);
2426	MMX_OP2(paddusw);
2427	MMX_OP2(pmaxub);
2428	MMX_OP2(pandn);
2429
2430	MMX_OP2(pmulhuw);
2431	MMX_OP2(pmulhw);
2432
2433	MMX_OP2(psubsb);
2434	MMX_OP2(psubsw);
2435	MMX_OP2(pminsw);
2436	MMX_OP2(por);
2437	MMX_OP2(paddsb);
2438	MMX_OP2(paddsw);
2439	MMX_OP2(pmaxsw);
2440	MMX_OP2(pxor);
2441	MMX_OP2(pmuludq);
2442	MMX_OP2(pmaddwd);
2443	MMX_OP2(psadbw);
2444	MMX_OP2(psubb);
2445	MMX_OP2(psubw);
2446	MMX_OP2(psubd);
2447	MMX_OP2(psubq);
2448	MMX_OP2(paddb);
2449	MMX_OP2(paddw);
2450	MMX_OP2(paddd);
2451
2452	MMX_OP2(pavgb);
2453	MMX_OP2(pavgw);
2454
2455	asm volatile ("pinsrw $1, %1, %0" : "=y" (r.q[0]) : "r" (0x12345678));
2456	printf("%-9s: r=" FMT64X "\n", "pinsrw", r.q[0]);
2457
2458	asm volatile ("pinsrw $5, %1, %0" : "=x" (r.dq) : "r" (0x12345678));
2459	printf("%-9s: r=" FMT64X "" FMT64X "\n", "pinsrw", r.q[1], r.q[0]);
2460
2461	a.q[0] = test_values[0][0];
2462	a.q[1] = test_values[0][1];
2463	asm volatile ("pextrw $1, %1, %0" : "=r" (r.l[0]) : "y" (a.q[0]));
2464	printf("%-9s: r=%08x\n", "pextrw", r.l[0]);
2465
2466	asm volatile ("pextrw $5, %1, %0" : "=r" (r.l[0]) : "x" (a.dq));
2467	printf("%-9s: r=%08x\n", "pextrw", r.l[0]);
2468
2469	asm volatile ("pmovmskb %1, %0" : "=r" (r.l[0]) : "y" (a.q[0]));
2470	printf("%-9s: r=%08x\n", "pmovmskb", r.l[0]);
2471
2472	asm volatile ("pmovmskb %1, %0" : "=r" (r.l[0]) : "x" (a.dq));
2473	printf("%-9s: r=%08x\n", "pmovmskb", r.l[0]);
2474
2475	{
2476	r.q[0] = -1;
2477	r.q[1] = -1;
2478
2479	a.q[0] = test_values[0][0];
2480	a.q[1] = test_values[0][1];
2481	b.q[0] = test_values[1][0];
2482	b.q[1] = test_values[1][1];
2483	asm volatile("maskmovq %1, %0" :
2484	: "y" (a.q[0]), "y" (b.q[0]), "D" (&r)
2485	: "memory");
2486	printf("%-9s: r=" FMT64X " a=" FMT64X " b=" FMT64X "\n",
2487	"maskmov",
2488	r.q[0],
2489	a.q[0],
2490	b.q[0]);
2491	asm volatile("maskmovdqu %1, %0" :
2492	: "x" (a.dq), "x" (b.dq), "D" (&r)
2493	: "memory");
2494	printf("%-9s: r=" FMT64X "" FMT64X " a=" FMT64X "" FMT64X " b=" FMT64X "" FMT64X "\n",
2495	"maskmov",
2496	r.q[1], r.q[0],
2497	a.q[1], a.q[0],
2498	b.q[1], b.q[0]);
2499	}
2500
2501	asm volatile ("emms");
2502
2503	SSE_OP2(punpcklqdq);
2504	SSE_OP2(punpckhqdq);
2505	SSE_OP2(andps);
2506	SSE_OP2(andpd);
2507	SSE_OP2(andnps);
2508	SSE_OP2(andnpd);
2509	SSE_OP2(orps);
2510	SSE_OP2(orpd);
2511	SSE_OP2(xorps);
2512	SSE_OP2(xorpd);
2513
2514	SSE_OP2(unpcklps);
2515	SSE_OP2(unpcklpd);
2516	SSE_OP2(unpckhps);
2517	SSE_OP2(unpckhpd);
2518
2519	SHUF_OP(shufps, 0x78);
2520	SHUF_OP(shufpd, 0x02);
2521
2522	PSHUF_OP(pshufd, 0x78);
2523	PSHUF_OP(pshuflw, 0x78);
2524	PSHUF_OP(pshufhw, 0x78);
2525
2526	SHIFT_OP(psrlw, 7);
2527	SHIFT_OP(psrlw, 16);
2528	SHIFT_OP(psraw, 7);
2529	SHIFT_OP(psraw, 16);
2530	SHIFT_OP(psllw, 7);
2531	SHIFT_OP(psllw, 16);
2532
2533	SHIFT_OP(psrld, 7);
2534	SHIFT_OP(psrld, 32);
2535	SHIFT_OP(psrad, 7);
2536	SHIFT_OP(psrad, 32);
2537	SHIFT_OP(pslld, 7);
2538	SHIFT_OP(pslld, 32);
2539
2540	SHIFT_OP(psrlq, 7);
2541	SHIFT_OP(psrlq, 32);
2542	SHIFT_OP(psllq, 7);
2543	SHIFT_OP(psllq, 32);
2544
2545	SHIFT_IM(psrldq, 16);
2546	SHIFT_IM(psrldq, 7);
2547	SHIFT_IM(pslldq, 16);
2548	SHIFT_IM(pslldq, 7);
2549
2550	MOVMSK(movmskps);
2551	MOVMSK(movmskpd);
2552
2553	/* FPU specific ops */
2554
2555	{
2556	uint32_t mxcsr;
2557	asm volatile("stmxcsr %0" : "=m" (mxcsr));
2558	printf("mxcsr=%08x\n", mxcsr & 0x1f80);
2559	asm volatile("ldmxcsr %0" : : "m" (mxcsr));
2560	}
2561
2562	test_sse_comi(2, -1);
2563	test_sse_comi(2, 2);
2564	test_sse_comi(2, 3);
2565	test_sse_comi(2, q_nan.d);
2566	test_sse_comi(q_nan.d, -1);
2567
2568	for(i = 0; i < 2; i++) {
2569	a.s[0] = 2.7;
2570	a.s[1] = 3.4;
2571	a.s[2] = 4;
2572	a.s[3] = -6.3;
2573	b.s[0] = 45.7;
2574	b.s[1] = 353.4;
2575	b.s[2] = 4;
2576	b.s[3] = 56.3;
2577	if (i == 1) {
2578	a.s[0] = q_nan.d;
2579	b.s[3] = q_nan.d;
2580	}
2581
2582	SSE_OPS(add);
2583	SSE_OPS(mul);
2584	SSE_OPS(sub);
2585	SSE_OPS(min);
2586	SSE_OPS(div);
2587	SSE_OPS(max);
2588	SSE_OPS(sqrt);
2589	SSE_OPS(cmpeq);
2590	SSE_OPS(cmplt);
2591	SSE_OPS(cmple);
2592	SSE_OPS(cmpunord);
2593	SSE_OPS(cmpneq);
2594	SSE_OPS(cmpnlt);
2595	SSE_OPS(cmpnle);
2596	SSE_OPS(cmpord);
2597
2598
2599	a.d[0] = 2.7;
2600	a.d[1] = -3.4;
2601	b.d[0] = 45.7;
2602	b.d[1] = -53.4;
2603	if (i == 1) {
2604	a.d[0] = q_nan.d;
2605	b.d[1] = q_nan.d;
2606	}
2607	SSE_OPD(add);
2608	SSE_OPD(mul);
2609	SSE_OPD(sub);
2610	SSE_OPD(min);
2611	SSE_OPD(div);
2612	SSE_OPD(max);
2613	SSE_OPD(sqrt);
2614	SSE_OPD(cmpeq);
2615	SSE_OPD(cmplt);
2616	SSE_OPD(cmple);
2617	SSE_OPD(cmpunord);
2618	SSE_OPD(cmpneq);
2619	SSE_OPD(cmpnlt);
2620	SSE_OPD(cmpnle);
2621	SSE_OPD(cmpord);
2622	}
2623
2624	/* float to float/int */
2625	a.s[0] = 2.7;
2626	a.s[1] = 3.4;
2627	a.s[2] = 4;
2628	a.s[3] = -6.3;
2629	CVT_OP_XMM(cvtps2pd);
2630	CVT_OP_XMM(cvtss2sd);
2631	CVT_OP_XMM2MMX(cvtps2pi);
2632	CVT_OP_XMM2MMX(cvttps2pi);
2633	CVT_OP_XMM2REG(cvtss2si);
2634	CVT_OP_XMM2REG(cvttss2si);
2635	CVT_OP_XMM(cvtps2dq);
2636	CVT_OP_XMM(cvttps2dq);
2637
2638	a.d[0] = 2.6;
2639	a.d[1] = -3.4;
2640	CVT_OP_XMM(cvtpd2ps);
2641	CVT_OP_XMM(cvtsd2ss);
2642	CVT_OP_XMM2MMX(cvtpd2pi);
2643	CVT_OP_XMM2MMX(cvttpd2pi);
2644	CVT_OP_XMM2REG(cvtsd2si);
2645	CVT_OP_XMM2REG(cvttsd2si);
2646	CVT_OP_XMM(cvtpd2dq);
2647	CVT_OP_XMM(cvttpd2dq);
2648
2649	/* sse/mmx moves */
2650	CVT_OP_XMM2MMX(movdq2q);
2651	CVT_OP_MMX2XMM(movq2dq);
2652
2653	/* int to float */
2654	a.l[0] = -6;
2655	a.l[1] = 2;
2656	a.l[2] = 100;
2657	a.l[3] = -60000;
2658	CVT_OP_MMX2XMM(cvtpi2ps);
2659	CVT_OP_MMX2XMM(cvtpi2pd);
2660	CVT_OP_REG2XMM(cvtsi2ss);
2661	CVT_OP_REG2XMM(cvtsi2sd);
2662	CVT_OP_XMM(cvtdq2ps);
2663	CVT_OP_XMM(cvtdq2pd);
2664
2665	/* XXX: test PNI insns */
2666	#if 0
2667	SSE_OP2(movshdup);
2668	#endif
2669	asm volatile ("emms");
2670	}
2671
2672	#endif
2673
2674	#define TEST_CONV_RAX(op)\
2675	{\
2676	unsigned long a, r;\
2677	a = i2l(0x8234a6f8);\
2678	r = a;\
2679	asm volatile(#op : "=a" (r) : "0" (r));\
2680	printf("%-10s A=" FMTLX " R=" FMTLX "\n", #op, a, r);\
2681	}
2682
2683	#define TEST_CONV_RAX_RDX(op)\
2684	{\
2685	unsigned long a, d, r, rh; \
2686	a = i2l(0x8234a6f8);\
2687	d = i2l(0x8345a1f2);\
2688	r = a;\
2689	rh = d;\
2690	asm volatile(#op : "=a" (r), "=d" (rh) : "0" (r), "1" (rh)); \
2691	printf("%-10s A=" FMTLX " R=" FMTLX ":" FMTLX "\n", #op, a, r, rh); \
2692	}
2693
2694	void test_conv(void)
2695	{
2696	TEST_CONV_RAX(cbw);
2697	TEST_CONV_RAX(cwde);
2698	#if defined(__x86_64__)
2699	TEST_CONV_RAX(cdqe);
2700	#endif
2701
2702	TEST_CONV_RAX_RDX(cwd);
2703	TEST_CONV_RAX_RDX(cdq);
2704	#if defined(__x86_64__)
2705	TEST_CONV_RAX_RDX(cqo);
2706	#endif
2707
2708	{
2709	unsigned long a, r;
2710	a = i2l(0x12345678);
2711	asm volatile("bswapl %k0" : "=r" (r) : "0" (a));
2712	printf("%-10s: A=" FMTLX " R=" FMTLX "\n", "bswapl", a, r);
2713	}
2714	#if defined(__x86_64__)
2715	{
2716	unsigned long a, r;
2717	a = i2l(0x12345678);
2718	asm volatile("bswapq %0" : "=r" (r) : "0" (a));
2719	printf("%-10s: A=" FMTLX " R=" FMTLX "\n", "bswapq", a, r);
2720	}
2721	#endif
2722	}
2723
2724	extern void *__start_initcall;
2725	extern void *__stop_initcall;
2726
2727
2728	int main(int argc, char **argv)
2729	{
2730	void **ptr;
2731	void (*func)(void);
2732
2733	ptr = &__start_initcall;
2734	while (ptr != &__stop_initcall) {
2735	func = *ptr++;
2736	func();
2737	}
2738	test_bsx();
2739	test_mul();
2740	test_jcc();
2741	test_loop();
2742	test_floats();
2743	#if !defined(__x86_64__)
2744	test_bcd();
2745	#endif
2746	test_xchg();
2747	test_string();
2748	test_misc();
2749	test_lea();
2750	#ifdef TEST_SEGS
2751	test_segs();
2752	test_code16();
2753	#endif
2754	#ifdef TEST_VM86
2755	test_vm86();
2756	#endif
2757	#if !defined(__x86_64__)
2758	test_exceptions();
2759	test_self_modifying_code();
2760	test_single_step();
2761	#endif
2762	test_enter();
2763	test_conv();
2764	#ifdef TEST_SSE
2765	test_sse();
2766	test_fxsave();
2767	#endif
2768	return 0;
2769	}

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

source: vbox/trunk/src/recompiler/tests/test-i386.c@ 77978

Download in other formats: