1 /* $Id: common_x86.c,v 1.14 2001/03/28 20:44:43 gareth Exp $ */
4 * Mesa 3-D graphics library
7 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
9 * Permission is hereby granted, free of charge, to any person obtaining a
10 * copy of this software and associated documentation files (the "Software"),
11 * to deal in the Software without restriction, including without limitation
12 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
13 * and/or sell copies of the Software, and to permit persons to whom the
14 * Software is furnished to do so, subject to the following conditions:
16 * The above copyright notice and this permission notice shall be included
17 * in all copies or substantial portions of the Software.
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
23 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
24 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
28 * Check CPU capabilities & initialize optimized funtions for this particular
31 * Written by Holger Waechtler <holger@akaflieg.extern.tu-berlin.de>
32 * Changed by Andre Werthmann <wertmann@cs.uni-potsdam.de> for using the
33 * new Katmai functions.
38 #if defined(USE_KATMAI_ASM) && defined(__linux__)
43 #include "common_x86_asm.h"
46 int _mesa_x86_cpu_features
= 0;
48 /* No reason for this to be public.
50 extern int _mesa_identify_x86_cpu_features( void );
53 static void message( const char *msg
)
59 if ( getenv( "MESA_DEBUG" ) ) {
66 fprintf( stderr
, "%s", msg
);
70 #if defined(USE_KATMAI_ASM)
72 * We must verify that the Streaming SIMD Extensions are truly supported
73 * on this processor before we go ahead and hook out the optimized code.
74 * Unfortunately, the CPUID bit isn't enough, as the OS must set the
75 * OSFXSR bit in CR4 if it supports the extended FPU save and restore
76 * required to use SSE. Unfortunately, we can't just go ahead and read
77 * this register, as only the kernel can do that. Similarly, we must
78 * verify that the OSXMMEXCPT bit in CR4 has been set by the OS,
79 * signifying that it supports unmasked SIMD FPU exceptions. If we take
80 * an unmasked exception and the OS doesn't correctly support them, the
81 * best we'll get is a SIGILL and the worst we'll get is an infinite
82 * loop in the signal delivery from the kernel as we can't interact with
83 * the SIMD FPU state to clear the exception bits. Either way, this is
87 extern void _mesa_test_os_katmai_support( void );
88 extern void _mesa_test_os_katmai_exception_support( void );
90 #if defined(__linux__) && defined(_POSIX_SOURCE)
91 static void sigill_handler( int signal
, struct sigcontext sc
)
93 message( "SIGILL, " );
95 /* Both the "xorps %%xmm0,%%xmm0" and "divps %xmm0,%%xmm1"
96 * instructions are 3 bytes long. We must increment the instruction
97 * pointer manually to avoid repeated execution of the offending
100 * If the SIGILL is caused by a divide-by-zero when unmasked
101 * exceptions aren't supported, the SIMD FPU status and control
102 * word will be restored at the end of the test, so we don't need
103 * to worry about doing it here. Besides, we may not be able to...
107 _mesa_x86_cpu_features
&= ~(X86_FEATURE_XMM
);
110 static void sigfpe_handler( int signal
, struct sigcontext sc
)
112 message( "SIGFPE, " );
114 if ( sc
.fpstate
->magic
!= 0xffff ) {
115 /* Our signal context has the extended FPU state, so reset the
116 * divide-by-zero exception mask and clear the divide-by-zero
119 sc
.fpstate
->mxcsr
|= 0x00000200;
120 sc
.fpstate
->mxcsr
&= 0xfffffffb;
122 /* If we ever get here, we're completely hosed.
125 _mesa_problem( NULL
, "SSE enabling test failed badly!" );
128 #endif /* __linux__ && _POSIX_SOURCE */
130 /* If we're running on a processor that can do SSE, let's see if we
131 * are allowed to or not. This will catch 2.4.0 or later kernels that
132 * haven't been configured for a Pentium III but are running on one,
133 * and RedHat patched 2.2 kernels that have broken exception handling
134 * support for user space apps that do SSE.
136 * GH: Isn't this just awful?
138 static void check_os_katmai_support( void )
140 #if defined(__linux__)
141 #if defined(_POSIX_SOURCE)
142 struct sigaction saved_sigill
;
143 struct sigaction saved_sigfpe
;
145 /* Save the original signal handlers.
147 sigaction( SIGILL
, NULL
, &saved_sigill
);
148 sigaction( SIGFPE
, NULL
, &saved_sigfpe
);
150 signal( SIGILL
, (void (*)(int))sigill_handler
);
151 signal( SIGFPE
, (void (*)(int))sigfpe_handler
);
153 /* Emulate test for OSFXSR in CR4. The OS will set this bit if it
154 * supports the extended FPU save and restore required for SSE. If
155 * we execute an SSE instruction on a PIII and get a SIGILL, the OS
156 * doesn't support Streaming SIMD Exceptions, even if the processor
160 message( "Testing OS support for SSE... " );
162 _mesa_test_os_katmai_support();
171 /* Emulate test for OSXMMEXCPT in CR4. The OS will set this bit if
172 * it supports unmasked SIMD FPU exceptions. If we unmask the
173 * exceptions, do a SIMD divide-by-zero and get a SIGILL, the OS
174 * doesn't support unmasked SIMD FPU exceptions. If we get a SIGFPE
175 * as expected, we're okay but we need to clean up after it.
177 * Are we being too stringent in our requirement that the OS support
178 * unmasked exceptions? Certain RedHat 2.2 kernels enable SSE by
179 * setting CR4.OSFXSR but don't support unmasked exceptions. Win98
180 * doesn't even support them. We at least know the user-space SSE
181 * support is good in kernels that do support unmasked exceptions,
182 * and therefore to be safe I'm going to leave this test in here.
185 message( "Testing OS support for SSE unmasked exceptions... " );
187 _mesa_test_os_katmai_exception_support();
196 /* Restore the original signal handlers.
198 sigaction( SIGILL
, &saved_sigill
, NULL
);
199 sigaction( SIGFPE
, &saved_sigfpe
, NULL
);
201 /* If we've gotten to here and the XMM CPUID bit is still set, we're
202 * safe to go ahead and hook out the SSE code throughout Mesa.
205 message( "Tests of OS support for SSE passed.\n" );
207 message( "Tests of OS support for SSE failed!\n" );
210 /* We can't use POSIX signal handling to test the availability of
211 * SSE, so we disable it by default.
213 message( "Cannot test OS support for SSE, disabling to be safe.\n" );
214 _mesa_x86_cpu_features
&= ~(X86_FEATURE_XMM
);
215 #endif /* _POSIX_SOURCE */
217 /* Do nothing on non-Linux platforms for now.
219 message( "Not testing OS support for SSE, leaving enabled.\n" );
220 #endif /* __linux__ */
223 #endif /* USE_KATMAI_ASM */
226 void _mesa_init_all_x86_transform_asm( void )
229 _mesa_x86_cpu_features
= _mesa_identify_x86_cpu_features();
231 if ( getenv( "MESA_NO_ASM" ) ) {
232 _mesa_x86_cpu_features
= 0;
235 if ( _mesa_x86_cpu_features
) {
236 _mesa_init_x86_transform_asm();
241 if ( getenv( "MESA_NO_MMX" ) == 0 ) {
242 message( "MMX cpu detected.\n" );
244 _mesa_x86_cpu_features
&= ~(X86_FEATURE_MMX
);
250 if ( cpu_has_3dnow
) {
251 if ( getenv( "MESA_NO_3DNOW" ) == 0 ) {
252 message( "3DNow! cpu detected.\n" );
253 _mesa_init_3dnow_transform_asm();
255 _mesa_x86_cpu_features
&= ~(X86_FEATURE_3DNOW
);
260 #ifdef USE_KATMAI_ASM
261 if ( cpu_has_xmm
&& getenv( "MESA_FORCE_KATMAI" ) == 0 ) {
262 check_os_katmai_support();
265 if ( getenv( "MESA_NO_KATMAI" ) == 0 ) {
266 message( "Katmai cpu detected.\n" );
267 _mesa_init_katmai_transform_asm();
269 _mesa_x86_cpu_features
&= ~(X86_FEATURE_XMM
);
276 /* Note: the above function must be called before this one, so that
277 * _mesa_x86_cpu_features gets correctly initialized.
279 void _mesa_init_all_x86_vertex_asm( void )
282 if ( _mesa_x86_cpu_features
) {
283 _mesa_init_x86_vertex_asm();
287 if ( cpu_has_3dnow
&& getenv( "MESA_NO_3DNOW" ) == 0 ) {
288 _mesa_init_3dnow_vertex_asm();
292 #ifdef USE_KATMAI_ASM
293 if ( cpu_has_xmm
&& getenv( "MESA_NO_KATMAI" ) == 0 ) {
294 _mesa_init_katmai_vertex_asm();