Rev 7164 | Rev 7168 | 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: 7165 $ |
||
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 |
||
7165 | clevermous | 37 | cmp ebx, fpu_data_size |
38 | ja $ |
||
7124 | dunkaist | 39 | |
40 | test eax, XCR0_AVX512 |
||
41 | jz @f |
||
42 | call init_avx512 |
||
7165 | clevermous | 43 | xsave [fpu_data] |
7124 | dunkaist | 44 | ret |
45 | @@: |
||
46 | test eax, XCR0_AVX |
||
47 | jz @f |
||
48 | call init_avx |
||
7165 | clevermous | 49 | xsave [fpu_data] |
7124 | dunkaist | 50 | ret |
51 | @@: |
||
52 | test eax, XCR0_SSE |
||
7164 | clevermous | 53 | jnz .sse |
54 | jmp .fpu_mmx |
||
7124 | dunkaist | 55 | .no_xsave: |
56 | mov [xsave_area_size], 512 ; enough for FPU/MMX and SSE |
||
2288 | clevermous | 57 | bt [cpu_caps], CAPS_SSE |
7164 | clevermous | 58 | jnc .fpu_mmx |
59 | .sse: |
||
7124 | dunkaist | 60 | call init_sse |
61 | fxsave [fpu_data] |
||
62 | ret |
||
7164 | clevermous | 63 | .fpu_mmx: |
7124 | dunkaist | 64 | call init_fpu_mmx |
65 | fnsave [fpu_data] |
||
66 | ret |
||
2288 | clevermous | 67 | |
7124 | dunkaist | 68 | init_fpu_mmx: |
69 | mov ecx, cr0 |
||
70 | and ecx, not CR0_EM |
||
71 | or ecx, CR0_MP + CR0_NE |
||
72 | mov cr0, ecx |
||
73 | ret |
||
74 | |||
75 | init_sse: |
||
2288 | clevermous | 76 | mov ebx, cr4 |
77 | mov ecx, cr0 |
||
78 | or ebx, CR4_OSFXSR+CR4_OSXMMEXPT |
||
79 | mov cr4, ebx |
||
80 | |||
7124 | dunkaist | 81 | and ecx, not (CR0_EM + CR0_MP) |
2288 | clevermous | 82 | or ecx, CR0_NE |
83 | mov cr0, ecx |
||
84 | |||
7124 | dunkaist | 85 | mov dword [esp-4], MXCSR_INIT |
2288 | clevermous | 86 | ldmxcsr [esp-4] |
87 | |||
88 | xorps xmm0, xmm0 |
||
89 | xorps xmm1, xmm1 |
||
90 | xorps xmm2, xmm2 |
||
91 | xorps xmm3, xmm3 |
||
92 | xorps xmm4, xmm4 |
||
93 | xorps xmm5, xmm5 |
||
94 | xorps xmm6, xmm6 |
||
95 | xorps xmm7, xmm7 |
||
96 | ret |
||
7124 | dunkaist | 97 | |
98 | init_avx: |
||
99 | mov ebx, cr4 |
||
100 | or ebx, CR4_OSFXSR + CR4_OSXMMEXPT |
||
101 | mov cr4, ebx |
||
102 | |||
2288 | clevermous | 103 | mov ecx, cr0 |
7124 | dunkaist | 104 | and ecx, not (CR0_EM + CR0_MP) |
105 | or ecx, CR0_NE |
||
2288 | clevermous | 106 | mov cr0, ecx |
7124 | dunkaist | 107 | |
108 | mov dword [esp-4], MXCSR_INIT |
||
109 | vldmxcsr [esp-4] |
||
110 | |||
111 | vzeroall |
||
2288 | clevermous | 112 | ret |
113 | |||
7124 | dunkaist | 114 | init_avx512: |
115 | mov ebx, cr4 |
||
116 | or ebx, CR4_OSFXSR + CR4_OSXMMEXPT |
||
117 | mov cr4, ebx |
||
118 | |||
119 | mov ecx, cr0 |
||
120 | and ecx, not (CR0_EM + CR0_MP) |
||
121 | or ecx, CR0_NE |
||
122 | mov cr0, ecx |
||
123 | |||
124 | mov dword [esp-4], MXCSR_INIT |
||
125 | vldmxcsr [esp-4] |
||
126 | |||
127 | vpxorq zmm0, zmm0, zmm0 |
||
128 | vpxorq zmm1, zmm1, zmm1 |
||
129 | vpxorq zmm2, zmm2, zmm2 |
||
130 | vpxorq zmm3, zmm3, zmm3 |
||
131 | vpxorq zmm4, zmm4, zmm4 |
||
132 | vpxorq zmm5, zmm5, zmm5 |
||
133 | vpxorq zmm6, zmm6, zmm6 |
||
134 | vpxorq zmm7, zmm7, zmm7 |
||
135 | |||
136 | ret |
||
137 | |||
2288 | clevermous | 138 | ; param |
139 | ; eax= 512 bytes memory area |
||
140 | |||
141 | align 4 |
||
142 | fpu_save: |
||
143 | push ecx |
||
144 | push esi |
||
145 | push edi |
||
146 | |||
147 | pushfd |
||
148 | cli |
||
149 | |||
150 | clts |
||
151 | mov edi, eax |
||
152 | |||
153 | mov ecx, [fpu_owner] |
||
154 | mov esi, [CURRENT_TASK] |
||
155 | cmp ecx, esi |
||
156 | jne .save |
||
157 | |||
158 | call save_context |
||
159 | jmp .exit |
||
160 | .save: |
||
161 | mov [fpu_owner], esi |
||
162 | |||
163 | shl ecx, 8 |
||
164 | mov eax, [ecx+SLOT_BASE+APPDATA.fpu_state] |
||
165 | |||
166 | call save_context |
||
167 | |||
168 | shl esi, 8 |
||
169 | mov esi, [esi+SLOT_BASE+APPDATA.fpu_state] |
||
170 | mov ecx, 512/4 |
||
171 | cld |
||
172 | rep movsd |
||
173 | fninit |
||
174 | .exit: |
||
175 | popfd |
||
176 | pop edi |
||
177 | pop esi |
||
178 | pop ecx |
||
179 | ret |
||
180 | |||
181 | align 4 |
||
182 | save_context: |
||
7124 | dunkaist | 183 | bt [cpu_caps+(CAPS_OSXSAVE/8)], CAPS_OSXSAVE mod 8 |
184 | jnc .no_xsave |
||
185 | xsave [eax] |
||
186 | ret |
||
187 | .no_xsave: |
||
2288 | clevermous | 188 | bt [cpu_caps], CAPS_SSE |
189 | jnc .no_SSE |
||
190 | |||
191 | fxsave [eax] |
||
192 | ret |
||
193 | .no_SSE: |
||
194 | fnsave [eax] |
||
195 | ret |
||
196 | |||
197 | align 4 |
||
198 | fpu_restore: |
||
199 | push ecx |
||
200 | push esi |
||
201 | |||
202 | mov esi, eax |
||
203 | |||
204 | pushfd |
||
205 | cli |
||
206 | |||
207 | mov ecx, [fpu_owner] |
||
208 | mov eax, [CURRENT_TASK] |
||
209 | cmp ecx, eax |
||
210 | jne .copy |
||
211 | |||
212 | clts |
||
7124 | dunkaist | 213 | bt [cpu_caps+(CAPS_OSXSAVE/8)], CAPS_OSXSAVE mod 8 |
214 | jnc .no_xsave |
||
215 | xrstor [esi] |
||
216 | popfd |
||
217 | pop esi |
||
218 | pop ecx |
||
219 | ret |
||
220 | .no_xsave: |
||
2288 | clevermous | 221 | bt [cpu_caps], CAPS_SSE |
222 | jnc .no_SSE |
||
223 | |||
224 | fxrstor [esi] |
||
225 | popfd |
||
226 | pop esi |
||
227 | pop ecx |
||
228 | ret |
||
229 | .no_SSE: |
||
230 | fnclex ;fix possible problems |
||
231 | frstor [esi] |
||
232 | popfd |
||
233 | pop esi |
||
234 | pop ecx |
||
235 | ret |
||
236 | .copy: |
||
237 | shl eax, 8 |
||
238 | mov edi, [eax+SLOT_BASE+APPDATA.fpu_state] |
||
239 | mov ecx, 512/4 |
||
240 | cld |
||
241 | rep movsd |
||
242 | popfd |
||
243 | pop esi |
||
244 | pop ecx |
||
245 | ret |
||
246 | |||
247 | align 4 |
||
248 | except_7: ;#NM exception handler |
||
249 | save_ring3_context |
||
250 | clts |
||
251 | mov ax, app_data; |
||
252 | mov ds, ax |
||
253 | mov es, ax |
||
254 | |||
255 | mov ebx, [fpu_owner] |
||
256 | cmp ebx, [CURRENT_TASK] |
||
257 | je .exit |
||
258 | |||
259 | shl ebx, 8 |
||
260 | mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state] |
||
261 | bt [cpu_caps], CAPS_SSE |
||
262 | jnc .no_SSE |
||
263 | |||
264 | fxsave [eax] |
||
265 | mov ebx, [CURRENT_TASK] |
||
266 | mov [fpu_owner], ebx |
||
267 | shl ebx, 8 |
||
268 | mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state] |
||
269 | fxrstor [eax] |
||
270 | .exit: |
||
271 | restore_ring3_context |
||
272 | iret |
||
273 | |||
274 | .no_SSE: |
||
275 | fnsave [eax] |
||
276 | mov ebx, [CURRENT_TASK] |
||
277 | mov [fpu_owner], ebx |
||
278 | shl ebx, 8 |
||
279 | mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state] |
||
280 | frstor [eax] |
||
281 | restore_ring3_context |
||
282 | iret |
||
283 | |||
284 | iglobal |
||
3534 | clevermous | 285 | fpu_owner dd 2 |
2288 | clevermous | 286 | endg |