Rev 5363 | Rev 7164 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
2288 | clevermous | 1 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
2 | ;; ;; |
||
7124 | dunkaist | 3 | ;; Copyright (C) KolibriOS team 2004-2017. All rights reserved. ;; |
2288 | clevermous | 4 | ;; Distributed under terms of the GNU General Public License ;; |
5 | ;; ;; |
||
6 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
||
7 | |||
8 | $Revision: 7124 $ |
||
9 | |||
10 | |||
11 | init_fpu: |
||
12 | clts |
||
13 | fninit |
||
14 | |||
7124 | dunkaist | 15 | bt [cpu_caps+(CAPS_XSAVE/8)], CAPS_XSAVE mod 8 |
16 | jnc .no_xsave |
||
17 | |||
18 | mov ecx, cr4 |
||
19 | or ecx, CR4_OSXSAVE |
||
20 | mov cr4, ecx |
||
21 | |||
22 | mov eax, 0x0d |
||
23 | xor ecx, ecx |
||
24 | cpuid |
||
25 | mov ebx, XCR0_FPU_MMX + XCR0_SSE + XCR0_AVX + XCR0_AVX512 |
||
26 | and ebx, eax |
||
27 | xor ecx, ecx |
||
28 | xgetbv |
||
29 | or eax, ebx |
||
30 | xor ecx, ecx |
||
31 | xsetbv |
||
32 | |||
33 | mov eax, 0x0d |
||
34 | xor ecx, ecx |
||
35 | cpuid |
||
36 | mov [xsave_area_size], ebx |
||
37 | |||
38 | test eax, XCR0_AVX512 |
||
39 | jz @f |
||
40 | call init_avx512 |
||
41 | ret |
||
42 | @@: |
||
43 | test eax, XCR0_AVX |
||
44 | jz @f |
||
45 | call init_avx |
||
46 | ret |
||
47 | @@: |
||
48 | test eax, XCR0_SSE |
||
49 | jz @f |
||
50 | call init_sse |
||
51 | ret |
||
52 | @@: |
||
53 | call init_fpu_mmx |
||
54 | ret |
||
55 | .no_xsave: |
||
56 | mov [xsave_area_size], 512 ; enough for FPU/MMX and SSE |
||
2288 | clevermous | 57 | bt [cpu_caps], CAPS_SSE |
7124 | dunkaist | 58 | jnc @f |
59 | call init_sse |
||
60 | fxsave [fpu_data] |
||
61 | ret |
||
62 | @@: |
||
63 | call init_fpu_mmx |
||
64 | fnsave [fpu_data] |
||
65 | ret |
||
2288 | clevermous | 66 | |
7124 | dunkaist | 67 | init_fpu_mmx: |
68 | mov ecx, cr0 |
||
69 | and ecx, not CR0_EM |
||
70 | or ecx, CR0_MP + CR0_NE |
||
71 | mov cr0, ecx |
||
72 | ret |
||
73 | |||
74 | init_sse: |
||
2288 | clevermous | 75 | mov ebx, cr4 |
76 | mov ecx, cr0 |
||
77 | or ebx, CR4_OSFXSR+CR4_OSXMMEXPT |
||
78 | mov cr4, ebx |
||
79 | |||
7124 | dunkaist | 80 | and ecx, not (CR0_EM + CR0_MP) |
2288 | clevermous | 81 | or ecx, CR0_NE |
82 | mov cr0, ecx |
||
83 | |||
7124 | dunkaist | 84 | mov dword [esp-4], MXCSR_INIT |
2288 | clevermous | 85 | ldmxcsr [esp-4] |
86 | |||
87 | xorps xmm0, xmm0 |
||
88 | xorps xmm1, xmm1 |
||
89 | xorps xmm2, xmm2 |
||
90 | xorps xmm3, xmm3 |
||
91 | xorps xmm4, xmm4 |
||
92 | xorps xmm5, xmm5 |
||
93 | xorps xmm6, xmm6 |
||
94 | xorps xmm7, xmm7 |
||
95 | ret |
||
7124 | dunkaist | 96 | |
97 | init_avx: |
||
98 | mov ebx, cr4 |
||
99 | or ebx, CR4_OSFXSR + CR4_OSXMMEXPT |
||
100 | mov cr4, ebx |
||
101 | |||
2288 | clevermous | 102 | mov ecx, cr0 |
7124 | dunkaist | 103 | and ecx, not (CR0_EM + CR0_MP) |
104 | or ecx, CR0_NE |
||
2288 | clevermous | 105 | mov cr0, ecx |
7124 | dunkaist | 106 | |
107 | mov dword [esp-4], MXCSR_INIT |
||
108 | vldmxcsr [esp-4] |
||
109 | |||
110 | vzeroall |
||
2288 | clevermous | 111 | ret |
112 | |||
7124 | dunkaist | 113 | init_avx512: |
114 | mov ebx, cr4 |
||
115 | or ebx, CR4_OSFXSR + CR4_OSXMMEXPT |
||
116 | mov cr4, ebx |
||
117 | |||
118 | mov ecx, cr0 |
||
119 | and ecx, not (CR0_EM + CR0_MP) |
||
120 | or ecx, CR0_NE |
||
121 | mov cr0, ecx |
||
122 | |||
123 | mov dword [esp-4], MXCSR_INIT |
||
124 | vldmxcsr [esp-4] |
||
125 | |||
126 | vpxorq zmm0, zmm0, zmm0 |
||
127 | vpxorq zmm1, zmm1, zmm1 |
||
128 | vpxorq zmm2, zmm2, zmm2 |
||
129 | vpxorq zmm3, zmm3, zmm3 |
||
130 | vpxorq zmm4, zmm4, zmm4 |
||
131 | vpxorq zmm5, zmm5, zmm5 |
||
132 | vpxorq zmm6, zmm6, zmm6 |
||
133 | vpxorq zmm7, zmm7, zmm7 |
||
134 | |||
135 | ret |
||
136 | |||
2288 | clevermous | 137 | ; param |
138 | ; eax= 512 bytes memory area |
||
139 | |||
140 | align 4 |
||
141 | fpu_save: |
||
142 | push ecx |
||
143 | push esi |
||
144 | push edi |
||
145 | |||
146 | pushfd |
||
147 | cli |
||
148 | |||
149 | clts |
||
150 | mov edi, eax |
||
151 | |||
152 | mov ecx, [fpu_owner] |
||
153 | mov esi, [CURRENT_TASK] |
||
154 | cmp ecx, esi |
||
155 | jne .save |
||
156 | |||
157 | call save_context |
||
158 | jmp .exit |
||
159 | .save: |
||
160 | mov [fpu_owner], esi |
||
161 | |||
162 | shl ecx, 8 |
||
163 | mov eax, [ecx+SLOT_BASE+APPDATA.fpu_state] |
||
164 | |||
165 | call save_context |
||
166 | |||
167 | shl esi, 8 |
||
168 | mov esi, [esi+SLOT_BASE+APPDATA.fpu_state] |
||
169 | mov ecx, 512/4 |
||
170 | cld |
||
171 | rep movsd |
||
172 | fninit |
||
173 | .exit: |
||
174 | popfd |
||
175 | pop edi |
||
176 | pop esi |
||
177 | pop ecx |
||
178 | ret |
||
179 | |||
180 | align 4 |
||
181 | save_context: |
||
7124 | dunkaist | 182 | bt [cpu_caps+(CAPS_OSXSAVE/8)], CAPS_OSXSAVE mod 8 |
183 | jnc .no_xsave |
||
184 | xsave [eax] |
||
185 | ret |
||
186 | .no_xsave: |
||
2288 | clevermous | 187 | bt [cpu_caps], CAPS_SSE |
188 | jnc .no_SSE |
||
189 | |||
190 | fxsave [eax] |
||
191 | ret |
||
192 | .no_SSE: |
||
193 | fnsave [eax] |
||
194 | ret |
||
195 | |||
196 | align 4 |
||
197 | fpu_restore: |
||
198 | push ecx |
||
199 | push esi |
||
200 | |||
201 | mov esi, eax |
||
202 | |||
203 | pushfd |
||
204 | cli |
||
205 | |||
206 | mov ecx, [fpu_owner] |
||
207 | mov eax, [CURRENT_TASK] |
||
208 | cmp ecx, eax |
||
209 | jne .copy |
||
210 | |||
211 | clts |
||
7124 | dunkaist | 212 | bt [cpu_caps+(CAPS_OSXSAVE/8)], CAPS_OSXSAVE mod 8 |
213 | jnc .no_xsave |
||
214 | xrstor [esi] |
||
215 | popfd |
||
216 | pop esi |
||
217 | pop ecx |
||
218 | ret |
||
219 | .no_xsave: |
||
2288 | clevermous | 220 | bt [cpu_caps], CAPS_SSE |
221 | jnc .no_SSE |
||
222 | |||
223 | fxrstor [esi] |
||
224 | popfd |
||
225 | pop esi |
||
226 | pop ecx |
||
227 | ret |
||
228 | .no_SSE: |
||
229 | fnclex ;fix possible problems |
||
230 | frstor [esi] |
||
231 | popfd |
||
232 | pop esi |
||
233 | pop ecx |
||
234 | ret |
||
235 | .copy: |
||
236 | shl eax, 8 |
||
237 | mov edi, [eax+SLOT_BASE+APPDATA.fpu_state] |
||
238 | mov ecx, 512/4 |
||
239 | cld |
||
240 | rep movsd |
||
241 | popfd |
||
242 | pop esi |
||
243 | pop ecx |
||
244 | ret |
||
245 | |||
246 | align 4 |
||
247 | except_7: ;#NM exception handler |
||
248 | save_ring3_context |
||
249 | clts |
||
250 | mov ax, app_data; |
||
251 | mov ds, ax |
||
252 | mov es, ax |
||
253 | |||
254 | mov ebx, [fpu_owner] |
||
255 | cmp ebx, [CURRENT_TASK] |
||
256 | je .exit |
||
257 | |||
258 | shl ebx, 8 |
||
259 | mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state] |
||
260 | bt [cpu_caps], CAPS_SSE |
||
261 | jnc .no_SSE |
||
262 | |||
263 | fxsave [eax] |
||
264 | mov ebx, [CURRENT_TASK] |
||
265 | mov [fpu_owner], ebx |
||
266 | shl ebx, 8 |
||
267 | mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state] |
||
268 | fxrstor [eax] |
||
269 | .exit: |
||
270 | restore_ring3_context |
||
271 | iret |
||
272 | |||
273 | .no_SSE: |
||
274 | fnsave [eax] |
||
275 | mov ebx, [CURRENT_TASK] |
||
276 | mov [fpu_owner], ebx |
||
277 | shl ebx, 8 |
||
278 | mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state] |
||
279 | frstor [eax] |
||
280 | restore_ring3_context |
||
281 | iret |
||
282 | |||
283 | iglobal |
||
3534 | clevermous | 284 | fpu_owner dd 2 |
2288 | clevermous | 285 | endg |