WebSVN – Kolibri OS – Diff – /kernel/trunk/core/fpu.inc


;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;                                                              ;;
;; Copyright (C) KolibriOS team 2004-2017. All rights reserved. ;;
;; Distributed under terms of the GNU General Public License    ;;
;;                                                              ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
$Revision: 7199 $
 
 
init_fpu:
        clts
        fninit
 
        bt      [cpu_caps+(CAPS_XSAVE/32)*4], CAPS_XSAVE mod 32
        jmp     .no_xsave       ; not ready to be jnc so far
 
        mov     ecx, cr4
        or      ecx, CR4_OSXSAVE
        mov     cr4, ecx
 
        mov     eax, 0x0d
        xor     ecx, ecx
        cpuid
        mov     ebx, XCR0_FPU_MMX + XCR0_SSE + XCR0_AVX + XCR0_AVX512
        and     ebx, eax
        xor     ecx, ecx
        xgetbv
        or      eax, ebx
        xor     ecx, ecx
        xsetbv
 
        mov     eax, 0x0d
        xor     ecx, ecx
        cpuid
        mov     [xsave_area_size], ebx
        cmp     ebx, fpu_data_size
        ja      $
 
        test    eax, XCR0_AVX512
        jz      @f
        call    init_avx512
        xsave   [fpu_data]
        ret
@@:
        test    eax, XCR0_AVX
        jz      @f
        call    init_avx
        xsave   [fpu_data]
        ret
@@:
        test    eax, XCR0_SSE
        jnz     .sse
        jmp     .fpu_mmx
.no_xsave:
        mov     [xsave_area_size], 512  ; enough for FPU/MMX and SSE
        bt      [cpu_caps], CAPS_SSE
        jnc     .fpu_mmx
.sse:
        call    init_sse
        fxsave  [fpu_data]
        ret
.fpu_mmx:
        call    init_fpu_mmx
        fnsave  [fpu_data]
        ret
 
init_fpu_mmx:
        mov     ecx, cr0
        and     ecx, not CR0_EM
        or      ecx, CR0_MP + CR0_NE
        mov     cr0, ecx
        ret
 
init_sse:
        mov     ebx, cr4
        mov     ecx, cr0
        or      ebx, CR4_OSFXSR+CR4_OSXMMEXPT
        mov     cr4, ebx
 
        and     ecx, not (CR0_EM + CR0_MP)
        or      ecx, CR0_NE
        mov     cr0, ecx
 
        mov     dword [esp-4], MXCSR_INIT
        ldmxcsr [esp-4]
 
        xorps   xmm0, xmm0
        xorps   xmm1, xmm1
        xorps   xmm2, xmm2
        xorps   xmm3, xmm3
        xorps   xmm4, xmm4
        xorps   xmm5, xmm5
        xorps   xmm6, xmm6
        xorps   xmm7, xmm7
        ret
 
init_avx:
        mov     ebx, cr4
        or      ebx, CR4_OSFXSR + CR4_OSXMMEXPT
        mov     cr4, ebx
 
        mov     ecx, cr0
        and     ecx, not (CR0_EM + CR0_MP)
        or      ecx, CR0_NE
        mov     cr0, ecx
 
        mov     dword [esp-4], MXCSR_INIT
        vldmxcsr [esp-4]
 
        vzeroall
        ret
 
init_avx512:
        mov     ebx, cr4
        or      ebx, CR4_OSFXSR + CR4_OSXMMEXPT
        mov     cr4, ebx
 
        mov     ecx, cr0
        and     ecx, not (CR0_EM + CR0_MP)
        or      ecx, CR0_NE
        mov     cr0, ecx
 
        mov     dword [esp-4], MXCSR_INIT
        vldmxcsr [esp-4]
 
        vpxorq  zmm0, zmm0, zmm0
        vpxorq  zmm1, zmm1, zmm1
        vpxorq  zmm2, zmm2, zmm2
        vpxorq  zmm3, zmm3, zmm3
        vpxorq  zmm4, zmm4, zmm4
        vpxorq  zmm5, zmm5, zmm5
        vpxorq  zmm6, zmm6, zmm6
        vpxorq  zmm7, zmm7, zmm7
 
        ret
 
; param
;  eax= 512 bytes memory area aligned on a 16-byte boundary
 
align 4
fpu_save:
        push    ecx
        push    esi
        push    edi
 
        pushfd
        cli
 
        clts
        mov     edi, eax
 
        mov     ecx, [fpu_owner]
        mov     esi, [CURRENT_TASK]
        cmp     ecx, esi
        jne     .save
 
        call    save_fpu_context
        jmp     .exit
.save:
        mov     [fpu_owner], esi
 
        shl     ecx, 8
        mov     eax, [ecx+SLOT_BASE+APPDATA.fpu_state]
 
        call    save_context
 
; first 512 bytes of XSAVE area have the same format as FXSAVE
        shl     esi, 8
        mov     esi, [esi+SLOT_BASE+APPDATA.fpu_state]
        mov     ecx, 512/4
        cld
        rep movsd
        fninit
.exit:
        popfd
        pop     edi
        pop     esi
        pop     ecx
        ret
 
avx_save_size:
        mov     eax, [xsave_area_size]
        ret
 
; param
;  eax= avx_save_size() bytes memory area aligned on a 64-byte boundary
 
avx_save:
        push    ecx
        push    esi
        push    edi
 
        pushfd
        cli
 
        clts
        mov     edi, eax
 
        mov     ecx, [fpu_owner]
        mov     esi, [CURRENT_TASK]
        cmp     ecx, esi
        jne     .save
 
        call    save_context
        jmp     .exit
.save:
        mov     [fpu_owner], esi
 
        shl     ecx, 8
        mov     eax, [ecx+SLOT_BASE+APPDATA.fpu_state]
 
        call    save_context
 
        shl     esi, 8
        mov     esi, [esi+SLOT_BASE+APPDATA.fpu_state]
        mov     ecx, [xsave_area_size]
        add     ecx, 3
        shr     ecx, 2
        rep movsd
        fninit
.exit:
        popfd
        pop     edi
        pop     esi
        pop     ecx
        ret
 
align 4
save_context:
        bt      [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
        jnc     save_fpu_context
        xsave   [eax]
        ret
save_fpu_context:
        bt      [cpu_caps], CAPS_SSE
        jnc     .no_SSE
        fxsave  [eax]
        ret
.no_SSE:
        fnsave  [eax]
        ret
 
 
align 4
fpu_restore:
        push    ecx
        push    esi
 
        mov     esi, eax
 
        pushfd
        cli
 
        mov     ecx, [fpu_owner]
        mov     eax, [CURRENT_TASK]
        cmp     ecx, eax
        jne     .copy
 
        clts
        bt      [cpu_caps], CAPS_SSE
        jnc     .no_SSE
 
        fxrstor [esi]
        popfd
        pop     esi
        pop     ecx
        ret
.no_SSE:
        fnclex                  ;fix possible problems
        frstor  [esi]
        popfd
        pop     esi
        pop     ecx
        ret
.copy:
        shl     eax, 8
        mov     edi, [eax+SLOT_BASE+APPDATA.fpu_state]
        mov     ecx, 512/4
        cld
        rep movsd
        popfd
        pop     esi
        pop     ecx
        ret
 
avx_restore:
        push    ecx
        push    esi
 
        mov     esi, eax
 
        pushfd
        cli
 
        mov     ecx, [fpu_owner]
        mov     eax, [CURRENT_TASK]
        cmp     ecx, eax
        jne     .copy
 
        clts
        bt      [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
        jnc     .no_xsave
        xrstor  [esi]
        popfd
        pop     esi
        pop     ecx
        ret
.no_xsave:
        bt      [cpu_caps], CAPS_SSE
        jnc     .no_SSE
 
        fxrstor [esi]
        popfd
        pop     esi
        pop     ecx
        ret
.no_SSE:
        fnclex                  ;fix possible problems
        frstor  [esi]
        popfd
        pop     esi
        pop     ecx
        ret
.copy:
        shl     eax, 8
        mov     edi, [eax+SLOT_BASE+APPDATA.fpu_state]
        mov     ecx, [xsave_area_size]
        add     ecx, 3
        shr     ecx, 2
        cld
        rep movsd
        popfd
        pop     esi
        pop     ecx
        ret
 
align 4
except_7:                  ;#NM exception handler
        save_ring3_context
        clts
        mov     ax, app_data;
        mov     ds, ax
        mov     es, ax
 
        mov     ebx, [fpu_owner]
        cmp     ebx, [CURRENT_TASK]
        je      .exit
 
        shl     ebx, 8
        mov     eax, [ebx+SLOT_BASE+APPDATA.fpu_state]
        bt      [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
        jnc     .no_xsave
        xsave   [eax]
        mov     ebx, [CURRENT_TASK]
        mov     [fpu_owner], ebx
        shl     ebx, 8
        mov     eax, [ebx+SLOT_BASE+APPDATA.fpu_state]
        xrstor  [eax]
.exit:
        restore_ring3_context
        iret
.no_xsave:
        bt      [cpu_caps], CAPS_SSE
        jnc     .no_SSE
 
        fxsave  [eax]
        mov     ebx, [CURRENT_TASK]
        mov     [fpu_owner], ebx
        shl     ebx, 8
        mov     eax, [ebx+SLOT_BASE+APPDATA.fpu_state]
        fxrstor [eax]
        restore_ring3_context
        iret
 
.no_SSE:
        fnsave  [eax]
        mov     ebx, [CURRENT_TASK]
        mov     [fpu_owner], ebx
        shl     ebx, 8
        mov     eax, [ebx+SLOT_BASE+APPDATA.fpu_state]
        frstor  [eax]
        restore_ring3_context
        iret
 
iglobal
  fpu_owner dd 2
endg

Rev 7168	Rev 7199
1	;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;	1	;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
2	;; ;;	2	;; ;;
3	;; Copyright (C) KolibriOS team 2004-2017. All rights reserved. ;;	3	;; Copyright (C) KolibriOS team 2004-2017. All rights reserved. ;;
4	;; Distributed under terms of the GNU General Public License ;;	4	;; Distributed under terms of the GNU General Public License ;;
5	;; ;;	5	;; ;;
6	;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;	6	;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
7		7
8	$Revision: 7168 $	8	$Revision: 7199 $
9		9
10		10
11	init_fpu:	11	init_fpu:
12	clts	12	clts
13	fninit	13	fninit
14		14
15	bt [cpu_caps+(CAPS_XSAVE/32)], CAPS_XSAVE mod 32	15	bt [cpu_caps+(CAPS_XSAVE/32)*4], CAPS_XSAVE mod 32
16	jnc .no_xsave	16	jmp .no_xsave ; not ready to be jnc so far
17		17
18	mov ecx, cr4	18	mov ecx, cr4
19	or ecx, CR4_OSXSAVE	19	or ecx, CR4_OSXSAVE
20	mov cr4, ecx	20	mov cr4, ecx
21		21
22	mov eax, 0x0d	22	mov eax, 0x0d
23	xor ecx, ecx	23	xor ecx, ecx
24	cpuid	24	cpuid
25	mov ebx, XCR0_FPU_MMX + XCR0_SSE + XCR0_AVX + XCR0_AVX512	25	mov ebx, XCR0_FPU_MMX + XCR0_SSE + XCR0_AVX + XCR0_AVX512
26	and ebx, eax	26	and ebx, eax
27	xor ecx, ecx	27	xor ecx, ecx
28	xgetbv	28	xgetbv
29	or eax, ebx	29	or eax, ebx
30	xor ecx, ecx	30	xor ecx, ecx
31	xsetbv	31	xsetbv
32		32
33	mov eax, 0x0d	33	mov eax, 0x0d
34	xor ecx, ecx	34	xor ecx, ecx
35	cpuid	35	cpuid
36	mov [xsave_area_size], ebx	36	mov [xsave_area_size], ebx
37	cmp ebx, fpu_data_size	37	cmp ebx, fpu_data_size
38	ja $	38	ja $
39		39
40	test eax, XCR0_AVX512	40	test eax, XCR0_AVX512
41	jz @f	41	jz @f
42	call init_avx512	42	call init_avx512
43	xsave [fpu_data]	43	xsave [fpu_data]
44	ret	44	ret
45	@@:	45	@@:
46	test eax, XCR0_AVX	46	test eax, XCR0_AVX
47	jz @f	47	jz @f
48	call init_avx	48	call init_avx
49	xsave [fpu_data]	49	xsave [fpu_data]
50	ret	50	ret
51	@@:	51	@@:
52	test eax, XCR0_SSE	52	test eax, XCR0_SSE
53	jnz .sse	53	jnz .sse
54	jmp .fpu_mmx	54	jmp .fpu_mmx
55	.no_xsave:	55	.no_xsave:
56	mov [xsave_area_size], 512 ; enough for FPU/MMX and SSE	56	mov [xsave_area_size], 512 ; enough for FPU/MMX and SSE
57	bt [cpu_caps], CAPS_SSE	57	bt [cpu_caps], CAPS_SSE
58	jnc .fpu_mmx	58	jnc .fpu_mmx
59	.sse:	59	.sse:
60	call init_sse	60	call init_sse
61	fxsave [fpu_data]	61	fxsave [fpu_data]
62	ret	62	ret
63	.fpu_mmx:	63	.fpu_mmx:
64	call init_fpu_mmx	64	call init_fpu_mmx
65	fnsave [fpu_data]	65	fnsave [fpu_data]
66	ret	66	ret
67		67
68	init_fpu_mmx:	68	init_fpu_mmx:
69	mov ecx, cr0	69	mov ecx, cr0
70	and ecx, not CR0_EM	70	and ecx, not CR0_EM
71	or ecx, CR0_MP + CR0_NE	71	or ecx, CR0_MP + CR0_NE
72	mov cr0, ecx	72	mov cr0, ecx
73	ret	73	ret
74		74
75	init_sse:	75	init_sse:
76	mov ebx, cr4	76	mov ebx, cr4
77	mov ecx, cr0	77	mov ecx, cr0
78	or ebx, CR4_OSFXSR+CR4_OSXMMEXPT	78	or ebx, CR4_OSFXSR+CR4_OSXMMEXPT
79	mov cr4, ebx	79	mov cr4, ebx
80		80
81	and ecx, not (CR0_EM + CR0_MP)	81	and ecx, not (CR0_EM + CR0_MP)
82	or ecx, CR0_NE	82	or ecx, CR0_NE
83	mov cr0, ecx	83	mov cr0, ecx
84		84
85	mov dword [esp-4], MXCSR_INIT	85	mov dword [esp-4], MXCSR_INIT
86	ldmxcsr [esp-4]	86	ldmxcsr [esp-4]
87		87
88	xorps xmm0, xmm0	88	xorps xmm0, xmm0
89	xorps xmm1, xmm1	89	xorps xmm1, xmm1
90	xorps xmm2, xmm2	90	xorps xmm2, xmm2
91	xorps xmm3, xmm3	91	xorps xmm3, xmm3
92	xorps xmm4, xmm4	92	xorps xmm4, xmm4
93	xorps xmm5, xmm5	93	xorps xmm5, xmm5
94	xorps xmm6, xmm6	94	xorps xmm6, xmm6
95	xorps xmm7, xmm7	95	xorps xmm7, xmm7
96	ret	96	ret
97		97
98	init_avx:	98	init_avx:
99	mov ebx, cr4	99	mov ebx, cr4
100	or ebx, CR4_OSFXSR + CR4_OSXMMEXPT	100	or ebx, CR4_OSFXSR + CR4_OSXMMEXPT
101	mov cr4, ebx	101	mov cr4, ebx
102		102
103	mov ecx, cr0	103	mov ecx, cr0
104	and ecx, not (CR0_EM + CR0_MP)	104	and ecx, not (CR0_EM + CR0_MP)
105	or ecx, CR0_NE	105	or ecx, CR0_NE
106	mov cr0, ecx	106	mov cr0, ecx
107		107
108	mov dword [esp-4], MXCSR_INIT	108	mov dword [esp-4], MXCSR_INIT
109	vldmxcsr [esp-4]	109	vldmxcsr [esp-4]
110		110
111	vzeroall	111	vzeroall
112	ret	112	ret
113		113
114	init_avx512:	114	init_avx512:
115	mov ebx, cr4	115	mov ebx, cr4
116	or ebx, CR4_OSFXSR + CR4_OSXMMEXPT	116	or ebx, CR4_OSFXSR + CR4_OSXMMEXPT
117	mov cr4, ebx	117	mov cr4, ebx
118		118
119	mov ecx, cr0	119	mov ecx, cr0
120	and ecx, not (CR0_EM + CR0_MP)	120	and ecx, not (CR0_EM + CR0_MP)
121	or ecx, CR0_NE	121	or ecx, CR0_NE
122	mov cr0, ecx	122	mov cr0, ecx
123		123
124	mov dword [esp-4], MXCSR_INIT	124	mov dword [esp-4], MXCSR_INIT
125	vldmxcsr [esp-4]	125	vldmxcsr [esp-4]
126		126
127	vpxorq zmm0, zmm0, zmm0	127	vpxorq zmm0, zmm0, zmm0
128	vpxorq zmm1, zmm1, zmm1	128	vpxorq zmm1, zmm1, zmm1
129	vpxorq zmm2, zmm2, zmm2	129	vpxorq zmm2, zmm2, zmm2
130	vpxorq zmm3, zmm3, zmm3	130	vpxorq zmm3, zmm3, zmm3
131	vpxorq zmm4, zmm4, zmm4	131	vpxorq zmm4, zmm4, zmm4
132	vpxorq zmm5, zmm5, zmm5	132	vpxorq zmm5, zmm5, zmm5
133	vpxorq zmm6, zmm6, zmm6	133	vpxorq zmm6, zmm6, zmm6
134	vpxorq zmm7, zmm7, zmm7	134	vpxorq zmm7, zmm7, zmm7
135		135
136	ret	136	ret
137		137
138	; param	138	; param
139	; eax= 512 bytes memory area aligned on a 16-byte boundary	139	; eax= 512 bytes memory area aligned on a 16-byte boundary
140		140
141	align 4	141	align 4
142	fpu_save:	142	fpu_save:
143	push ecx	143	push ecx
144	push esi	144	push esi
145	push edi	145	push edi
146		146
147	pushfd	147	pushfd
148	cli	148	cli
149		149
150	clts	150	clts
151	mov edi, eax	151	mov edi, eax
152		152
153	mov ecx, [fpu_owner]	153	mov ecx, [fpu_owner]
154	mov esi, [CURRENT_TASK]	154	mov esi, [CURRENT_TASK]
155	cmp ecx, esi	155	cmp ecx, esi
156	jne .save	156	jne .save
157		157
158	call save_fpu_context	158	call save_fpu_context
159	jmp .exit	159	jmp .exit
160	.save:	160	.save:
161	mov [fpu_owner], esi	161	mov [fpu_owner], esi
162		162
163	shl ecx, 8	163	shl ecx, 8
164	mov eax, [ecx+SLOT_BASE+APPDATA.fpu_state]	164	mov eax, [ecx+SLOT_BASE+APPDATA.fpu_state]
165		165
166	call save_context	166	call save_context
167		167
168	; first 512 bytes of XSAVE area have the same format as FXSAVE	168	; first 512 bytes of XSAVE area have the same format as FXSAVE
169	shl esi, 8	169	shl esi, 8
170	mov esi, [esi+SLOT_BASE+APPDATA.fpu_state]	170	mov esi, [esi+SLOT_BASE+APPDATA.fpu_state]
171	mov ecx, 512/4	171	mov ecx, 512/4
172	cld	172	cld
173	rep movsd	173	rep movsd
174	fninit	174	fninit
175	.exit:	175	.exit:
176	popfd	176	popfd
177	pop edi	177	pop edi
178	pop esi	178	pop esi
179	pop ecx	179	pop ecx
180	ret	180	ret
181		181
182	avx_save_size:	182	avx_save_size:
183	mov eax, [xsave_area_size]	183	mov eax, [xsave_area_size]
184	ret	184	ret
185		185
186	; param	186	; param
187	; eax= avx_save_size() bytes memory area aligned on a 64-byte boundary	187	; eax= avx_save_size() bytes memory area aligned on a 64-byte boundary
188		188
189	avx_save:	189	avx_save:
190	push ecx	190	push ecx
191	push esi	191	push esi
192	push edi	192	push edi
193		193
194	pushfd	194	pushfd
195	cli	195	cli
196		196
197	clts	197	clts
198	mov edi, eax	198	mov edi, eax
199		199
200	mov ecx, [fpu_owner]	200	mov ecx, [fpu_owner]
201	mov esi, [CURRENT_TASK]	201	mov esi, [CURRENT_TASK]
202	cmp ecx, esi	202	cmp ecx, esi
203	jne .save	203	jne .save
204		204
205	call save_context	205	call save_context
206	jmp .exit	206	jmp .exit
207	.save:	207	.save:
208	mov [fpu_owner], esi	208	mov [fpu_owner], esi
209		209
210	shl ecx, 8	210	shl ecx, 8
211	mov eax, [ecx+SLOT_BASE+APPDATA.fpu_state]	211	mov eax, [ecx+SLOT_BASE+APPDATA.fpu_state]
212		212
213	call save_context	213	call save_context
214		214
215	shl esi, 8	215	shl esi, 8
216	mov esi, [esi+SLOT_BASE+APPDATA.fpu_state]	216	mov esi, [esi+SLOT_BASE+APPDATA.fpu_state]
217	mov ecx, [xsave_area_size]	217	mov ecx, [xsave_area_size]
218	add ecx, 3	218	add ecx, 3
219	shr ecx, 2	219	shr ecx, 2
220	rep movsd	220	rep movsd
221	fninit	221	fninit
222	.exit:	222	.exit:
223	popfd	223	popfd
224	pop edi	224	pop edi
225	pop esi	225	pop esi
226	pop ecx	226	pop ecx
227	ret	227	ret
228		228
229	align 4	229	align 4
230	save_context:	230	save_context:
231	bt [cpu_caps+(CAPS_OSXSAVE/32)], CAPS_OSXSAVE mod 32	231	bt [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
232	jnc save_fpu_context	232	jnc save_fpu_context
233	xsave [eax]	233	xsave [eax]
234	ret	234	ret
235	save_fpu_context:	235	save_fpu_context:
236	bt [cpu_caps], CAPS_SSE	236	bt [cpu_caps], CAPS_SSE
237	jnc .no_SSE	237	jnc .no_SSE
238	fxsave [eax]	238	fxsave [eax]
239	ret	239	ret
240	.no_SSE:	240	.no_SSE:
241	fnsave [eax]	241	fnsave [eax]
242	ret	242	ret
243		243
244		244
245	align 4	245	align 4
246	fpu_restore:	246	fpu_restore:
247	push ecx	247	push ecx
248	push esi	248	push esi
249		249
250	mov esi, eax	250	mov esi, eax
251		251
252	pushfd	252	pushfd
253	cli	253	cli
254		254
255	mov ecx, [fpu_owner]	255	mov ecx, [fpu_owner]
256	mov eax, [CURRENT_TASK]	256	mov eax, [CURRENT_TASK]
257	cmp ecx, eax	257	cmp ecx, eax
258	jne .copy	258	jne .copy
259		259
260	clts	260	clts
261	bt [cpu_caps], CAPS_SSE	261	bt [cpu_caps], CAPS_SSE
262	jnc .no_SSE	262	jnc .no_SSE
263		263
264	fxrstor [esi]	264	fxrstor [esi]
265	popfd	265	popfd
266	pop esi	266	pop esi
267	pop ecx	267	pop ecx
268	ret	268	ret
269	.no_SSE:	269	.no_SSE:
270	fnclex ;fix possible problems	270	fnclex ;fix possible problems
271	frstor [esi]	271	frstor [esi]
272	popfd	272	popfd
273	pop esi	273	pop esi
274	pop ecx	274	pop ecx
275	ret	275	ret
276	.copy:	276	.copy:
277	shl eax, 8	277	shl eax, 8
278	mov edi, [eax+SLOT_BASE+APPDATA.fpu_state]	278	mov edi, [eax+SLOT_BASE+APPDATA.fpu_state]
279	mov ecx, 512/4	279	mov ecx, 512/4
280	cld	280	cld
281	rep movsd	281	rep movsd
282	popfd	282	popfd
283	pop esi	283	pop esi
284	pop ecx	284	pop ecx
285	ret	285	ret
286		286
287	avx_restore:	287	avx_restore:
288	push ecx	288	push ecx
289	push esi	289	push esi
290		290
291	mov esi, eax	291	mov esi, eax
292		292
293	pushfd	293	pushfd
294	cli	294	cli
295		295
296	mov ecx, [fpu_owner]	296	mov ecx, [fpu_owner]
297	mov eax, [CURRENT_TASK]	297	mov eax, [CURRENT_TASK]
298	cmp ecx, eax	298	cmp ecx, eax
299	jne .copy	299	jne .copy
300		300
301	clts	301	clts
302	bt [cpu_caps+(CAPS_OSXSAVE/32)], CAPS_OSXSAVE mod 32	302	bt [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
303	jnc .no_xsave	303	jnc .no_xsave
304	xrstor [esi]	304	xrstor [esi]
305	popfd	305	popfd
306	pop esi	306	pop esi
307	pop ecx	307	pop ecx
308	ret	308	ret
309	.no_xsave:	309	.no_xsave:
310	bt [cpu_caps], CAPS_SSE	310	bt [cpu_caps], CAPS_SSE
311	jnc .no_SSE	311	jnc .no_SSE
312		312
313	fxrstor [esi]	313	fxrstor [esi]
314	popfd	314	popfd
315	pop esi	315	pop esi
316	pop ecx	316	pop ecx
317	ret	317	ret
318	.no_SSE:	318	.no_SSE:
319	fnclex ;fix possible problems	319	fnclex ;fix possible problems
320	frstor [esi]	320	frstor [esi]
321	popfd	321	popfd
322	pop esi	322	pop esi
323	pop ecx	323	pop ecx
324	ret	324	ret
325	.copy:	325	.copy:
326	shl eax, 8	326	shl eax, 8
327	mov edi, [eax+SLOT_BASE+APPDATA.fpu_state]	327	mov edi, [eax+SLOT_BASE+APPDATA.fpu_state]
328	mov ecx, [xsave_area_size]	328	mov ecx, [xsave_area_size]
329	add ecx, 3	329	add ecx, 3
330	shr ecx, 2	330	shr ecx, 2
331	cld	331	cld
332	rep movsd	332	rep movsd
333	popfd	333	popfd
334	pop esi	334	pop esi
335	pop ecx	335	pop ecx
336	ret	336	ret
337		337
338	align 4	338	align 4
339	except_7: ;#NM exception handler	339	except_7: ;#NM exception handler
340	save_ring3_context	340	save_ring3_context
341	clts	341	clts
342	mov ax, app_data;	342	mov ax, app_data;
343	mov ds, ax	343	mov ds, ax
344	mov es, ax	344	mov es, ax
345		345
346	mov ebx, [fpu_owner]	346	mov ebx, [fpu_owner]
347	cmp ebx, [CURRENT_TASK]	347	cmp ebx, [CURRENT_TASK]
348	je .exit	348	je .exit
349		349
350	shl ebx, 8	350	shl ebx, 8
351	mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state]	351	mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state]
352	bt [cpu_caps+(CAPS_OSXSAVE/32)], CAPS_OSXSAVE mod 32	352	bt [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
353	jnc .no_xsave	353	jnc .no_xsave
354	xsave [eax]	354	xsave [eax]
355	mov ebx, [CURRENT_TASK]	355	mov ebx, [CURRENT_TASK]
356	mov [fpu_owner], ebx	356	mov [fpu_owner], ebx
357	shl ebx, 8	357	shl ebx, 8
358	mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state]	358	mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state]
359	xrstor [eax]	359	xrstor [eax]
360	.exit:	360	.exit:
361	restore_ring3_context	361	restore_ring3_context
362	iret	362	iret
363	.no_xsave:	363	.no_xsave:
364	bt [cpu_caps], CAPS_SSE	364	bt [cpu_caps], CAPS_SSE
365	jnc .no_SSE	365	jnc .no_SSE
366		366
367	fxsave [eax]	367	fxsave [eax]
368	mov ebx, [CURRENT_TASK]	368	mov ebx, [CURRENT_TASK]
369	mov [fpu_owner], ebx	369	mov [fpu_owner], ebx
370	shl ebx, 8	370	shl ebx, 8
371	mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state]	371	mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state]
372	fxrstor [eax]	372	fxrstor [eax]
373	restore_ring3_context	373	restore_ring3_context
374	iret	374	iret
375		375
376	.no_SSE:	376	.no_SSE:
377	fnsave [eax]	377	fnsave [eax]
378	mov ebx, [CURRENT_TASK]	378	mov ebx, [CURRENT_TASK]
379	mov [fpu_owner], ebx	379	mov [fpu_owner], ebx
380	shl ebx, 8	380	shl ebx, 8
381	mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state]	381	mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state]
382	frstor [eax]	382	frstor [eax]
383	restore_ring3_context	383	restore_ring3_context
384	iret	384	iret
385		385
386	iglobal	386	iglobal
387	fpu_owner dd 2	387	fpu_owner dd 2
388	endg	388	endg

Subversion Repositories Kolibri OS

(root)/kernel/trunk/core/fpu.inc – Rev 7168 → 7199