bring over build fixes from stable branch
[mesa.git] / src / mesa / x86 / common_x86.c
1 /*
2 * Mesa 3-D graphics library
3 * Version: 6.0.1
4 *
5 * Copyright (C) 1999-2004 Brian Paul All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
13 *
14 * The above copyright notice and this permission notice shall be included
15 * in all copies or substantial portions of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
21 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
22 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
23 */
24
25 /*
26 * Check CPU capabilities & initialize optimized funtions for this particular
27 * processor.
28 *
29 * Written by Holger Waechtler <holger@akaflieg.extern.tu-berlin.de>
30 * Changed by Andre Werthmann <wertmann@cs.uni-potsdam.de> for using the
31 * new SSE functions.
32 */
33
34 /* XXX these includes should probably go into imports.h or glheader.h */
35 #if defined(USE_SSE_ASM) && defined(__linux__)
36 #include <signal.h>
37 #endif
38 #if defined(USE_SSE_ASM) && defined(__FreeBSD__)
39 #include <sys/types.h>
40 #include <sys/sysctl.h>
41 #endif
42
43 #include "common_x86_asm.h"
44 #include "imports.h"
45
46
47 int _mesa_x86_cpu_features = 0;
48
49 /* No reason for this to be public.
50 */
51 extern GLuint _ASMAPI _mesa_x86_has_cpuid(void);
52 extern void _ASMAPI _mesa_x86_cpuid(GLuint op, GLuint *reg_eax, GLuint *reg_ebx, GLuint *reg_ecx, GLuint *reg_edx);
53 extern GLuint _ASMAPI _mesa_x86_cpuid_eax(GLuint op);
54 extern GLuint _ASMAPI _mesa_x86_cpuid_ebx(GLuint op);
55 extern GLuint _ASMAPI _mesa_x86_cpuid_ecx(GLuint op);
56 extern GLuint _ASMAPI _mesa_x86_cpuid_edx(GLuint op);
57
58 static void message( const char *msg )
59 {
60 GLboolean debug;
61 #ifdef DEBUG
62 debug = GL_TRUE;
63 #else
64 if ( _mesa_getenv( "MESA_DEBUG" ) ) {
65 debug = GL_TRUE;
66 } else {
67 debug = GL_FALSE;
68 }
69 #endif
70 if ( debug ) {
71 fprintf( stderr, "%s", msg );
72 }
73 }
74
75 #if defined(USE_SSE_ASM)
76 /*
77 * We must verify that the Streaming SIMD Extensions are truly supported
78 * on this processor before we go ahead and hook out the optimized code.
79 * Unfortunately, the CPUID bit isn't enough, as the OS must set the
80 * OSFXSR bit in CR4 if it supports the extended FPU save and restore
81 * required to use SSE. Unfortunately, we can't just go ahead and read
82 * this register, as only the kernel can do that. Similarly, we must
83 * verify that the OSXMMEXCPT bit in CR4 has been set by the OS,
84 * signifying that it supports unmasked SIMD FPU exceptions. If we take
85 * an unmasked exception and the OS doesn't correctly support them, the
86 * best we'll get is a SIGILL and the worst we'll get is an infinite
87 * loop in the signal delivery from the kernel as we can't interact with
88 * the SIMD FPU state to clear the exception bits. Either way, this is
89 * not good.
90 */
91
92 extern void _mesa_test_os_sse_support( void );
93 extern void _mesa_test_os_sse_exception_support( void );
94
95 #if defined(__linux__) && defined(_POSIX_SOURCE) && defined(X86_FXSR_MAGIC)
96 static void sigill_handler( int signal, struct sigcontext sc )
97 {
98 message( "SIGILL, " );
99
100 /* Both the "xorps %%xmm0,%%xmm0" and "divps %xmm0,%%xmm1"
101 * instructions are 3 bytes long. We must increment the instruction
102 * pointer manually to avoid repeated execution of the offending
103 * instruction.
104 *
105 * If the SIGILL is caused by a divide-by-zero when unmasked
106 * exceptions aren't supported, the SIMD FPU status and control
107 * word will be restored at the end of the test, so we don't need
108 * to worry about doing it here. Besides, we may not be able to...
109 */
110 sc.eip += 3;
111
112 _mesa_x86_cpu_features &= ~(X86_FEATURE_XMM);
113 }
114
115 static void sigfpe_handler( int signal, struct sigcontext sc )
116 {
117 message( "SIGFPE, " );
118
119 if ( sc.fpstate->magic != 0xffff ) {
120 /* Our signal context has the extended FPU state, so reset the
121 * divide-by-zero exception mask and clear the divide-by-zero
122 * exception bit.
123 */
124 sc.fpstate->mxcsr |= 0x00000200;
125 sc.fpstate->mxcsr &= 0xfffffffb;
126 } else {
127 /* If we ever get here, we're completely hosed.
128 */
129 message( "\n\n" );
130 _mesa_problem( NULL, "SSE enabling test failed badly!" );
131 }
132 }
133 #endif /* __linux__ && _POSIX_SOURCE && X86_FXSR_MAGIC */
134
135 /* If we're running on a processor that can do SSE, let's see if we
136 * are allowed to or not. This will catch 2.4.0 or later kernels that
137 * haven't been configured for a Pentium III but are running on one,
138 * and RedHat patched 2.2 kernels that have broken exception handling
139 * support for user space apps that do SSE.
140 *
141 * GH: Isn't this just awful?
142 */
143 static void check_os_sse_support( void )
144 {
145 #if defined(__linux__)
146 #if defined(_POSIX_SOURCE) && defined(X86_FXSR_MAGIC)
147 struct sigaction saved_sigill;
148 struct sigaction saved_sigfpe;
149
150 /* Save the original signal handlers.
151 */
152 sigaction( SIGILL, NULL, &saved_sigill );
153 sigaction( SIGFPE, NULL, &saved_sigfpe );
154
155 signal( SIGILL, (void (*)(int))sigill_handler );
156 signal( SIGFPE, (void (*)(int))sigfpe_handler );
157
158 /* Emulate test for OSFXSR in CR4. The OS will set this bit if it
159 * supports the extended FPU save and restore required for SSE. If
160 * we execute an SSE instruction on a PIII and get a SIGILL, the OS
161 * doesn't support Streaming SIMD Exceptions, even if the processor
162 * does.
163 */
164 if ( cpu_has_xmm ) {
165 message( "Testing OS support for SSE... " );
166
167 _mesa_test_os_sse_support();
168
169 if ( cpu_has_xmm ) {
170 message( "yes.\n" );
171 } else {
172 message( "no!\n" );
173 }
174 }
175
176 /* Emulate test for OSXMMEXCPT in CR4. The OS will set this bit if
177 * it supports unmasked SIMD FPU exceptions. If we unmask the
178 * exceptions, do a SIMD divide-by-zero and get a SIGILL, the OS
179 * doesn't support unmasked SIMD FPU exceptions. If we get a SIGFPE
180 * as expected, we're okay but we need to clean up after it.
181 *
182 * Are we being too stringent in our requirement that the OS support
183 * unmasked exceptions? Certain RedHat 2.2 kernels enable SSE by
184 * setting CR4.OSFXSR but don't support unmasked exceptions. Win98
185 * doesn't even support them. We at least know the user-space SSE
186 * support is good in kernels that do support unmasked exceptions,
187 * and therefore to be safe I'm going to leave this test in here.
188 */
189 if ( cpu_has_xmm ) {
190 message( "Testing OS support for SSE unmasked exceptions... " );
191
192 _mesa_test_os_sse_exception_support();
193
194 if ( cpu_has_xmm ) {
195 message( "yes.\n" );
196 } else {
197 message( "no!\n" );
198 }
199 }
200
201 /* Restore the original signal handlers.
202 */
203 sigaction( SIGILL, &saved_sigill, NULL );
204 sigaction( SIGFPE, &saved_sigfpe, NULL );
205
206 /* If we've gotten to here and the XMM CPUID bit is still set, we're
207 * safe to go ahead and hook out the SSE code throughout Mesa.
208 */
209 if ( cpu_has_xmm ) {
210 message( "Tests of OS support for SSE passed.\n" );
211 } else {
212 message( "Tests of OS support for SSE failed!\n" );
213 }
214 #else
215 /* We can't use POSIX signal handling to test the availability of
216 * SSE, so we disable it by default.
217 */
218 message( "Cannot test OS support for SSE, disabling to be safe.\n" );
219 _mesa_x86_cpu_features &= ~(X86_FEATURE_XMM);
220 #endif /* _POSIX_SOURCE && X86_FXSR_MAGIC */
221 #elif defined(__FreeBSD__)
222 {
223 int ret, len, enabled;
224 len = sizeof(enabled);
225 ret = sysctlbyname("hw.instruction_sse", &enabled, &len, NULL, 0);
226 if (ret || !enabled)
227 _mesa_x86_cpu_features &= ~(X86_FEATURE_XMM);
228 }
229 #else
230 /* Do nothing on other platforms for now.
231 */
232 message( "Not testing OS support for SSE, leaving enabled.\n" );
233 #endif /* __linux__ */
234 }
235
236 #endif /* USE_SSE_ASM */
237
238
239 void _mesa_init_all_x86_transform_asm( void )
240 {
241 (void) message; /* silence warning */
242 #ifdef USE_X86_ASM
243 _mesa_x86_cpu_features = 0;
244
245 if (!_mesa_x86_has_cpuid()) {
246 message("CPUID not detected");
247 }
248 else {
249 GLuint cpu_features;
250 GLuint cpu_ext_features;
251 GLuint cpu_ext_info;
252 char cpu_vendor[13];
253 GLuint result;
254
255 /* get vendor name */
256 _mesa_x86_cpuid(0, &result, (GLuint *)(cpu_vendor + 0), (GLuint *)(cpu_vendor + 8), (GLuint *)(cpu_vendor + 4));
257 cpu_vendor[12] = '\0';
258
259 message("cpu vendor: ");
260 message(cpu_vendor);
261 message("\n");
262
263 /* get cpu features */
264 cpu_features = _mesa_x86_cpuid_edx(1);
265
266 if (cpu_features & X86_CPU_FPU)
267 _mesa_x86_cpu_features |= X86_FEATURE_FPU;
268 if (cpu_features & X86_CPU_CMOV)
269 _mesa_x86_cpu_features |= X86_FEATURE_CMOV;
270
271 #ifdef USE_MMX_ASM
272 if (cpu_features & X86_CPU_MMX)
273 _mesa_x86_cpu_features |= X86_FEATURE_MMX;
274 #endif
275
276 #ifdef USE_SSE_ASM
277 if (cpu_features & X86_CPU_XMM)
278 _mesa_x86_cpu_features |= X86_FEATURE_XMM;
279 if (cpu_features & X86_CPU_XMM2)
280 _mesa_x86_cpu_features |= X86_FEATURE_XMM2;
281 #endif
282
283 /* query extended cpu features */
284 if ((cpu_ext_info = _mesa_x86_cpuid_eax(0x80000000)) > 0x80000000) {
285 if (cpu_ext_info >= 0x80000001) {
286
287 cpu_ext_features = _mesa_x86_cpuid_edx(0x80000001);
288
289 if (cpu_features & X86_CPU_MMX) {
290
291 #ifdef USE_3DNOW_ASM
292 if (cpu_ext_features & X86_CPUEXT_3DNOW)
293 _mesa_x86_cpu_features |= X86_FEATURE_3DNOW;
294 if (cpu_ext_features & X86_CPUEXT_3DNOW_EXT)
295 _mesa_x86_cpu_features |= X86_FEATURE_3DNOWEXT;
296 #endif
297
298 #ifdef USE_MMX_ASM
299 if (cpu_ext_features & X86_CPUEXT_MMX_EXT)
300 _mesa_x86_cpu_features |= X86_FEATURE_MMXEXT;
301 #endif
302 }
303 }
304
305 /* query cpu name */
306 if (cpu_ext_info >= 0x80000002) {
307 GLuint ofs;
308 char cpu_name[49];
309 for (ofs = 0; ofs < 3; ofs++)
310 _mesa_x86_cpuid(0x80000002+ofs, (GLuint *)(cpu_name + (16*ofs)+0), (GLuint *)(cpu_name + (16*ofs)+4), (GLuint *)(cpu_name + (16*ofs)+8), (GLuint *)(cpu_name + (16*ofs)+12));
311 cpu_name[48] = '\0'; /* the name should be NULL terminated, but just to be sure */
312
313 message("cpu name: ");
314 message(cpu_name);
315 message("\n");
316 }
317 }
318
319 }
320
321 if ( _mesa_getenv( "MESA_NO_ASM" ) ) {
322 _mesa_x86_cpu_features = 0;
323 }
324
325 if ( _mesa_x86_cpu_features ) {
326 _mesa_init_x86_transform_asm();
327 }
328
329 #ifdef USE_MMX_ASM
330 if ( cpu_has_mmx ) {
331 if ( _mesa_getenv( "MESA_NO_MMX" ) == 0 ) {
332 message( "MMX cpu detected.\n" );
333 } else {
334 _mesa_x86_cpu_features &= ~(X86_FEATURE_MMX);
335 }
336 }
337 #endif
338
339 #ifdef USE_3DNOW_ASM
340 if ( cpu_has_3dnow ) {
341 if ( _mesa_getenv( "MESA_NO_3DNOW" ) == 0 ) {
342 message( "3DNow! cpu detected.\n" );
343 _mesa_init_3dnow_transform_asm();
344 } else {
345 _mesa_x86_cpu_features &= ~(X86_FEATURE_3DNOW);
346 }
347 }
348 #endif
349
350 #ifdef USE_SSE_ASM
351 if ( cpu_has_xmm && _mesa_getenv( "MESA_FORCE_SSE" ) == 0 ) {
352 check_os_sse_support();
353 }
354 if ( cpu_has_xmm ) {
355 if (_mesa_getenv( "MESA_NO_SSE" ) == 0 ) {
356 message( "SSE cpu detected.\n" );
357 _mesa_init_sse_transform_asm();
358 } else {
359 message( "SSE cpu detected, but switched off by user.\n" );
360 _mesa_x86_cpu_features &= ~(X86_FEATURE_XMM);
361 }
362 }
363 #endif
364 #endif
365 }
366