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


;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;                                                              ;;
;; Copyright (C) KolibriOS team 2004-2022. All rights reserved. ;;
;; Distributed under terms of the GNU General Public License    ;;
;;                                                              ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
$Revision: 9715 $
 
 
init_fpu:
        clts
        fninit
 
        bt      [cpu_caps+(CAPS_XSAVE/32)*4], CAPS_XSAVE mod 32
        jnc     .no_xsave
 
        mov     ecx, cr4
        or      ecx, CR4_OSXSAVE
        mov     cr4, ecx
        ; don't call cpuid again
        bts     [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
 
        ; zero xsave header
        mov     ecx, 64/4
        xor     eax, eax
        mov     edi, fpu_data + 512     ; skip legacy region
        rep stosd
 
        mov     eax, 0x0d       ; extended state enumeration main leaf
        xor     ecx, ecx
        cpuid
        and     eax, XCR0_FPU_MMX + XCR0_SSE + XCR0_AVX + XCR0_AVX512
        xor     edx, edx
        mov     [xsave_eax], eax
        mov     [xsave_edx], edx
        xor     ecx, ecx
        xsetbv
 
        mov     eax, 0x0d
        xor     ecx, ecx
        cpuid
        add     ebx, 63
        and     ebx, NOT 63
        mov     [xsave_area_size], ebx
        cmp     ebx, fpu_data_size
        ja      $
 
        test    eax, XCR0_AVX512
        jz      @f
        call    init_avx512
        mov     eax, [xsave_eax]
        mov     edx, [xsave_edx]
        xsave   [fpu_data]
        ret
@@:
        test    eax, XCR0_AVX
        jz      @f
        call    init_avx
        mov     eax, [xsave_eax]
        mov     edx, [xsave_edx]
        xsave   [fpu_data]
        ret
@@:
        test    eax, XCR0_SSE
        jz      $
        call    init_sse
        mov     eax, [xsave_eax]
        mov     edx, [xsave_edx]
        xsave   [fpu_data]
        ret
.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_slot_idx]
        cmp     ecx, esi
        jne     .save
 
        call    save_fpu_context
        jmp     .exit
.save:
        mov     [fpu_owner], esi
 
        shl     ecx, BSF sizeof.APPDATA
        mov     eax, [SLOT_BASE + ecx + APPDATA.fpu_state]
 
        call    save_context
 
; first 512 bytes of XSAVE area have the same format as FXSAVE
        shl     esi, BSF sizeof.APPDATA
        mov     esi, [SLOT_BASE + esi + 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
 
align 4
avx_save:
        push    ecx
        push    esi
        push    edi
 
        pushfd
        cli
 
        clts
        mov     edi, eax
 
        mov     ecx, [fpu_owner]
        mov     esi, [current_slot_idx]
        cmp     ecx, esi
        jne     .save
 
        call    save_context
        jmp     .exit
.save:
        mov     [fpu_owner], esi
 
        shl     ecx, BSF sizeof.APPDATA
        mov     eax, [SLOT_BASE + ecx + APPDATA.fpu_state]
 
        call    save_context
 
        shl     esi, BSF sizeof.APPDATA
        mov     esi, [SLOT_BASE + esi + 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
        push    eax edx
        mov     ecx, eax
        mov     eax, [xsave_eax]
        mov     edx, [xsave_edx]
        xsave   [ecx]
        pop     edx 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_slot_idx]
        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, BSF sizeof.APPDATA
        mov     edi, [SLOT_BASE + eax + APPDATA.fpu_state]
        mov     ecx, 512/4
        cld
        rep movsd
        popfd
        pop     esi
        pop     ecx
        ret
 
align 4
avx_restore:
        push    ecx
        push    esi
 
        mov     esi, eax
 
        pushfd
        cli
 
        mov     ecx, [fpu_owner]
        mov     eax, [current_slot_idx]
        cmp     ecx, eax
        jne     .copy
 
        clts
        bt      [cpu_caps + (CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
        jnc     .no_xsave
        push    edx
        mov     eax, [xsave_eax]
        mov     edx, [xsave_edx]
        xrstor  [esi]
        pop     edx
        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, BSF sizeof.APPDATA
        mov     edi, [SLOT_BASE + eax + 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_slot_idx]
        je      .exit
 
        shl     ebx, BSF sizeof.APPDATA
        mov     eax, [SLOT_BASE + ebx + APPDATA.fpu_state]
        bt      [cpu_caps + (CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
        jnc     .no_xsave
        mov     ecx, eax
        mov     eax, [xsave_eax]
        mov     edx, [xsave_edx]
        xsave   [ecx]
        mov     ebx, [current_slot_idx]
        mov     [fpu_owner], ebx
        shl     ebx, BSF sizeof.APPDATA
        mov     ecx, [SLOT_BASE + ebx + APPDATA.fpu_state]
        xrstor  [ecx]
.exit:
        restore_ring3_context
        iret
.no_xsave:
        bt      [cpu_caps], CAPS_SSE
        jnc     .no_SSE
 
        fxsave  [eax]
        mov     ebx, [current_slot_idx]
        mov     [fpu_owner], ebx
        shl     ebx, BSF sizeof.APPDATA
        mov     eax, [SLOT_BASE + ebx + APPDATA.fpu_state]
        fxrstor [eax]
        restore_ring3_context
        iret
 
.no_SSE:
        fnsave  [eax]
        mov     ebx, [current_slot_idx]
        mov     [fpu_owner], ebx
        shl     ebx, BSF sizeof.APPDATA
        mov     eax, [SLOT_BASE + ebx + APPDATA.fpu_state]
        frstor  [eax]
        restore_ring3_context
        iret
 
iglobal
  fpu_owner dd 2
endg

Rev 8869	Rev 9715
1	;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;	1	;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
2	;; ;;	2	;; ;;
3	;; Copyright (C) KolibriOS team 2004-2017. All rights reserved. ;;	3	;; Copyright (C) KolibriOS team 2004-2022. 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: 8869 $	8	$Revision: 9715 $
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)*4], CAPS_XSAVE mod 32	15	bt [cpu_caps+(CAPS_XSAVE/32)*4], CAPS_XSAVE mod 32
16	jnc .no_xsave	16	jnc .no_xsave
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	; don't call cpuid again	21	; don't call cpuid again
22	bts [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32	22	bts [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
23		23
24	; zero xsave header	24	; zero xsave header
25	mov ecx, 64/4	25	mov ecx, 64/4
26	xor eax, eax	26	xor eax, eax
27	mov edi, fpu_data + 512 ; skip legacy region	27	mov edi, fpu_data + 512 ; skip legacy region
28	rep stosd	28	rep stosd
29		29
30	mov eax, 0x0d ; extended state enumeration main leaf	30	mov eax, 0x0d ; extended state enumeration main leaf
31	xor ecx, ecx	31	xor ecx, ecx
32	cpuid	32	cpuid
33	and eax, XCR0_FPU_MMX + XCR0_SSE + XCR0_AVX + XCR0_AVX512	33	and eax, XCR0_FPU_MMX + XCR0_SSE + XCR0_AVX + XCR0_AVX512
34	xor edx, edx	34	xor edx, edx
35	mov [xsave_eax], eax	35	mov [xsave_eax], eax
36	mov [xsave_edx], edx	36	mov [xsave_edx], edx
37	xor ecx, ecx	37	xor ecx, ecx
38	xsetbv	38	xsetbv
39		39
40	mov eax, 0x0d	40	mov eax, 0x0d
41	xor ecx, ecx	41	xor ecx, ecx
42	cpuid	42	cpuid
43	add ebx, 63	43	add ebx, 63
44	and ebx, NOT 63	44	and ebx, NOT 63
45	mov [xsave_area_size], ebx	45	mov [xsave_area_size], ebx
46	cmp ebx, fpu_data_size	46	cmp ebx, fpu_data_size
47	ja $	47	ja $
48		48
49	test eax, XCR0_AVX512	49	test eax, XCR0_AVX512
50	jz @f	50	jz @f
51	call init_avx512	51	call init_avx512
52	mov eax, [xsave_eax]	52	mov eax, [xsave_eax]
53	mov edx, [xsave_edx]	53	mov edx, [xsave_edx]
54	xsave [fpu_data]	54	xsave [fpu_data]
55	ret	55	ret
56	@@:	56	@@:
57	test eax, XCR0_AVX	57	test eax, XCR0_AVX
58	jz @f	58	jz @f
59	call init_avx	59	call init_avx
60	mov eax, [xsave_eax]	60	mov eax, [xsave_eax]
61	mov edx, [xsave_edx]	61	mov edx, [xsave_edx]
62	xsave [fpu_data]	62	xsave [fpu_data]
63	ret	63	ret
64	@@:	64	@@:
65	test eax, XCR0_SSE	65	test eax, XCR0_SSE
66	jz $	66	jz $
67	call init_sse	67	call init_sse
68	mov eax, [xsave_eax]	68	mov eax, [xsave_eax]
69	mov edx, [xsave_edx]	69	mov edx, [xsave_edx]
70	xsave [fpu_data]	70	xsave [fpu_data]
71	ret	71	ret
72	.no_xsave:	72	.no_xsave:
73	mov [xsave_area_size], 512 ; enough for FPU/MMX and SSE	73	mov [xsave_area_size], 512 ; enough for FPU/MMX and SSE
74	bt [cpu_caps], CAPS_SSE	74	bt [cpu_caps], CAPS_SSE
75	jnc .fpu_mmx	75	jnc .fpu_mmx
76	.sse:	76	.sse:
77	call init_sse	77	call init_sse
78	fxsave [fpu_data]	78	fxsave [fpu_data]
79	ret	79	ret
80	.fpu_mmx:	80	.fpu_mmx:
81	call init_fpu_mmx	81	call init_fpu_mmx
82	fnsave [fpu_data]	82	fnsave [fpu_data]
83	ret	83	ret
84		84
85	init_fpu_mmx:	85	init_fpu_mmx:
86	mov ecx, cr0	86	mov ecx, cr0
87	and ecx, not CR0_EM	87	and ecx, not CR0_EM
88	or ecx, CR0_MP + CR0_NE	88	or ecx, CR0_MP + CR0_NE
89	mov cr0, ecx	89	mov cr0, ecx
90	ret	90	ret
91		91
92	init_sse:	92	init_sse:
93	mov ebx, cr4	93	mov ebx, cr4
94	mov ecx, cr0	94	mov ecx, cr0
95	or ebx, CR4_OSFXSR + CR4_OSXMMEXPT	95	or ebx, CR4_OSFXSR + CR4_OSXMMEXPT
96	mov cr4, ebx	96	mov cr4, ebx
97		97
98	and ecx, not (CR0_EM + CR0_MP)	98	and ecx, not (CR0_EM + CR0_MP)
99	or ecx, CR0_NE	99	or ecx, CR0_NE
100	mov cr0, ecx	100	mov cr0, ecx
101		101
102	mov dword [esp-4], MXCSR_INIT	102	mov dword [esp-4], MXCSR_INIT
103	ldmxcsr [esp-4]	103	ldmxcsr [esp-4]
104		104
105	xorps xmm0, xmm0	105	xorps xmm0, xmm0
106	xorps xmm1, xmm1	106	xorps xmm1, xmm1
107	xorps xmm2, xmm2	107	xorps xmm2, xmm2
108	xorps xmm3, xmm3	108	xorps xmm3, xmm3
109	xorps xmm4, xmm4	109	xorps xmm4, xmm4
110	xorps xmm5, xmm5	110	xorps xmm5, xmm5
111	xorps xmm6, xmm6	111	xorps xmm6, xmm6
112	xorps xmm7, xmm7	112	xorps xmm7, xmm7
113	ret	113	ret
114		114
115	init_avx:	115	init_avx:
116	mov ebx, cr4	116	mov ebx, cr4
117	or ebx, CR4_OSFXSR + CR4_OSXMMEXPT	117	or ebx, CR4_OSFXSR + CR4_OSXMMEXPT
118	mov cr4, ebx	118	mov cr4, ebx
119		119
120	mov ecx, cr0	120	mov ecx, cr0
121	and ecx, not (CR0_EM + CR0_MP)	121	and ecx, not (CR0_EM + CR0_MP)
122	or ecx, CR0_NE	122	or ecx, CR0_NE
123	mov cr0, ecx	123	mov cr0, ecx
124		124
125	mov dword [esp-4], MXCSR_INIT	125	mov dword [esp-4], MXCSR_INIT
126	vldmxcsr [esp-4]	126	vldmxcsr [esp-4]
127		127
128	vzeroall	128	vzeroall
129	ret	129	ret
130		130
131	init_avx512:	131	init_avx512:
132	mov ebx, cr4	132	mov ebx, cr4
133	or ebx, CR4_OSFXSR + CR4_OSXMMEXPT	133	or ebx, CR4_OSFXSR + CR4_OSXMMEXPT
134	mov cr4, ebx	134	mov cr4, ebx
135		135
136	mov ecx, cr0	136	mov ecx, cr0
137	and ecx, not (CR0_EM + CR0_MP)	137	and ecx, not (CR0_EM + CR0_MP)
138	or ecx, CR0_NE	138	or ecx, CR0_NE
139	mov cr0, ecx	139	mov cr0, ecx
140		140
141	mov dword [esp-4], MXCSR_INIT	141	mov dword [esp-4], MXCSR_INIT
142	vldmxcsr [esp-4]	142	vldmxcsr [esp-4]
143		143
144	vpxorq zmm0, zmm0, zmm0	144	vpxorq zmm0, zmm0, zmm0
145	vpxorq zmm1, zmm1, zmm1	145	vpxorq zmm1, zmm1, zmm1
146	vpxorq zmm2, zmm2, zmm2	146	vpxorq zmm2, zmm2, zmm2
147	vpxorq zmm3, zmm3, zmm3	147	vpxorq zmm3, zmm3, zmm3
148	vpxorq zmm4, zmm4, zmm4	148	vpxorq zmm4, zmm4, zmm4
149	vpxorq zmm5, zmm5, zmm5	149	vpxorq zmm5, zmm5, zmm5
150	vpxorq zmm6, zmm6, zmm6	150	vpxorq zmm6, zmm6, zmm6
151	vpxorq zmm7, zmm7, zmm7	151	vpxorq zmm7, zmm7, zmm7
152		152
153	ret	153	ret
154		154
155	; param	155	; param
156	; eax= 512 bytes memory area aligned on a 16-byte boundary	156	; eax= 512 bytes memory area aligned on a 16-byte boundary
157		157
158	align 4	158	align 4
159	fpu_save:	159	fpu_save:
160	push ecx	160	push ecx
161	push esi	161	push esi
162	push edi	162	push edi
163		163
164	pushfd	164	pushfd
165	cli	165	cli
166		166
167	clts	167	clts
168	mov edi, eax	168	mov edi, eax
169		169
170	mov ecx, [fpu_owner]	170	mov ecx, [fpu_owner]
171	mov esi, [current_slot_idx]	171	mov esi, [current_slot_idx]
172	cmp ecx, esi	172	cmp ecx, esi
173	jne .save	173	jne .save
174		174
175	call save_fpu_context	175	call save_fpu_context
176	jmp .exit	176	jmp .exit
177	.save:	177	.save:
178	mov [fpu_owner], esi	178	mov [fpu_owner], esi
179		179
180	shl ecx, 8	180	shl ecx, BSF sizeof.APPDATA
181	mov eax, [ecx+SLOT_BASE+APPDATA.fpu_state]	181	mov eax, [SLOT_BASE + ecx + APPDATA.fpu_state]
182		182
183	call save_context	183	call save_context
184		184
185	; first 512 bytes of XSAVE area have the same format as FXSAVE	185	; first 512 bytes of XSAVE area have the same format as FXSAVE
186	shl esi, 8	186	shl esi, BSF sizeof.APPDATA
187	mov esi, [esi+SLOT_BASE+APPDATA.fpu_state]	187	mov esi, [SLOT_BASE + esi + APPDATA.fpu_state]
188	mov ecx, 512/4	188	mov ecx, 512/4
189	cld	189	cld
190	rep movsd	190	rep movsd
191	fninit	191	fninit
192	.exit:	192	.exit:
193	popfd	193	popfd
194	pop edi	194	pop edi
195	pop esi	195	pop esi
196	pop ecx	196	pop ecx
197	ret	197	ret
198		198
199	avx_save_size:	199	avx_save_size:
200	mov eax, [xsave_area_size]	200	mov eax, [xsave_area_size]
201	ret	201	ret
202		202
203	; param	203	; param
204	; eax= avx_save_size() bytes memory area aligned on a 64-byte boundary	204	; eax= avx_save_size() bytes memory area aligned on a 64-byte boundary
205		205
206	align 4	206	align 4
207	avx_save:	207	avx_save:
208	push ecx	208	push ecx
209	push esi	209	push esi
210	push edi	210	push edi
211		211
212	pushfd	212	pushfd
213	cli	213	cli
214		214
215	clts	215	clts
216	mov edi, eax	216	mov edi, eax
217		217
218	mov ecx, [fpu_owner]	218	mov ecx, [fpu_owner]
219	mov esi, [current_slot_idx]	219	mov esi, [current_slot_idx]
220	cmp ecx, esi	220	cmp ecx, esi
221	jne .save	221	jne .save
222		222
223	call save_context	223	call save_context
224	jmp .exit	224	jmp .exit
225	.save:	225	.save:
226	mov [fpu_owner], esi	226	mov [fpu_owner], esi
227		227
228	shl ecx, 8	228	shl ecx, BSF sizeof.APPDATA
229	mov eax, [ecx+SLOT_BASE+APPDATA.fpu_state]	229	mov eax, [SLOT_BASE + ecx + APPDATA.fpu_state]
230		230
231	call save_context	231	call save_context
232		232
233	shl esi, 8	233	shl esi, BSF sizeof.APPDATA
234	mov esi, [esi+SLOT_BASE+APPDATA.fpu_state]	234	mov esi, [SLOT_BASE + esi + APPDATA.fpu_state]
235	mov ecx, [xsave_area_size]	235	mov ecx, [xsave_area_size]
236	add ecx, 3	236	add ecx, 3
237	shr ecx, 2	237	shr ecx, 2
238	rep movsd	238	rep movsd
239	fninit	239	fninit
240	.exit:	240	.exit:
241	popfd	241	popfd
242	pop edi	242	pop edi
243	pop esi	243	pop esi
244	pop ecx	244	pop ecx
245	ret	245	ret
246		246
247	align 4	247	align 4
248	save_context:	248	save_context:
249	bt [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32	249	bt [cpu_caps + (CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
250	jnc save_fpu_context	250	jnc save_fpu_context
251	push eax edx	251	push eax edx
252	mov ecx, eax	252	mov ecx, eax
253	mov eax, [xsave_eax]	253	mov eax, [xsave_eax]
254	mov edx, [xsave_edx]	254	mov edx, [xsave_edx]
255	xsave [ecx]	255	xsave [ecx]
256	pop edx eax	256	pop edx eax
257	ret	257	ret
258	save_fpu_context:	258	save_fpu_context:
259	bt [cpu_caps], CAPS_SSE	259	bt [cpu_caps], CAPS_SSE
260	jnc .no_SSE	260	jnc .no_SSE
261	fxsave [eax]	261	fxsave [eax]
262	ret	262	ret
263	.no_SSE:	263	.no_SSE:
264	fnsave [eax]	264	fnsave [eax]
265	ret	265	ret
266		266
267		267
268	align 4	268	align 4
269	fpu_restore:	269	fpu_restore:
270	push ecx	270	push ecx
271	push esi	271	push esi
272		272
273	mov esi, eax	273	mov esi, eax
274		274
275	pushfd	275	pushfd
276	cli	276	cli
277		277
278	mov ecx, [fpu_owner]	278	mov ecx, [fpu_owner]
279	mov eax, [current_slot_idx]	279	mov eax, [current_slot_idx]
280	cmp ecx, eax	280	cmp ecx, eax
281	jne .copy	281	jne .copy
282		282
283	clts	283	clts
284	bt [cpu_caps], CAPS_SSE	284	bt [cpu_caps], CAPS_SSE
285	jnc .no_SSE	285	jnc .no_SSE
286		286
287	fxrstor [esi]	287	fxrstor [esi]
288	popfd	288	popfd
289	pop esi	289	pop esi
290	pop ecx	290	pop ecx
291	ret	291	ret
292	.no_SSE:	292	.no_SSE:
293	fnclex ;fix possible problems	293	fnclex ;fix possible problems
294	frstor [esi]	294	frstor [esi]
295	popfd	295	popfd
296	pop esi	296	pop esi
297	pop ecx	297	pop ecx
298	ret	298	ret
299	.copy:	299	.copy:
300	shl eax, 8	300	shl eax, BSF sizeof.APPDATA
301	mov edi, [eax+SLOT_BASE+APPDATA.fpu_state]	301	mov edi, [SLOT_BASE + eax + APPDATA.fpu_state]
302	mov ecx, 512/4	302	mov ecx, 512/4
303	cld	303	cld
304	rep movsd	304	rep movsd
305	popfd	305	popfd
306	pop esi	306	pop esi
307	pop ecx	307	pop ecx
308	ret	308	ret
309		309
310	align 4	310	align 4
311	avx_restore:	311	avx_restore:
312	push ecx	312	push ecx
313	push esi	313	push esi
314		314
315	mov esi, eax	315	mov esi, eax
316		316
317	pushfd	317	pushfd
318	cli	318	cli
319		319
320	mov ecx, [fpu_owner]	320	mov ecx, [fpu_owner]
321	mov eax, [current_slot_idx]	321	mov eax, [current_slot_idx]
322	cmp ecx, eax	322	cmp ecx, eax
323	jne .copy	323	jne .copy
324		324
325	clts	325	clts
326	bt [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32	326	bt [cpu_caps + (CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
327	jnc .no_xsave	327	jnc .no_xsave
328	push edx	328	push edx
329	mov eax, [xsave_eax]	329	mov eax, [xsave_eax]
330	mov edx, [xsave_edx]	330	mov edx, [xsave_edx]
331	xrstor [esi]	331	xrstor [esi]
332	pop edx	332	pop edx
333	popfd	333	popfd
334	pop esi	334	pop esi
335	pop ecx	335	pop ecx
336	ret	336	ret
337	.no_xsave:	337	.no_xsave:
338	bt [cpu_caps], CAPS_SSE	338	bt [cpu_caps], CAPS_SSE
339	jnc .no_SSE	339	jnc .no_SSE
340		340
341	fxrstor [esi]	341	fxrstor [esi]
342	popfd	342	popfd
343	pop esi	343	pop esi
344	pop ecx	344	pop ecx
345	ret	345	ret
346	.no_SSE:	346	.no_SSE:
347	fnclex ;fix possible problems	347	fnclex ;fix possible problems
348	frstor [esi]	348	frstor [esi]
349	popfd	349	popfd
350	pop esi	350	pop esi
351	pop ecx	351	pop ecx
352	ret	352	ret
353	.copy:	353	.copy:
354	shl eax, 8	354	shl eax, BSF sizeof.APPDATA
355	mov edi, [eax+SLOT_BASE+APPDATA.fpu_state]	355	mov edi, [SLOT_BASE + eax + APPDATA.fpu_state]
356	mov ecx, [xsave_area_size]	356	mov ecx, [xsave_area_size]
357	add ecx, 3	357	add ecx, 3
358	shr ecx, 2	358	shr ecx, 2
359	cld	359	cld
360	rep movsd	360	rep movsd
361	popfd	361	popfd
362	pop esi	362	pop esi
363	pop ecx	363	pop ecx
364	ret	364	ret
365		365
366	align 4	366	align 4
367	except_7: ;#NM exception handler	367	except_7: ;#NM exception handler
368	save_ring3_context	368	save_ring3_context
369	clts	369	clts
370	mov ax, app_data;	370	mov ax, app_data;
371	mov ds, ax	371	mov ds, ax
372	mov es, ax	372	mov es, ax
373		373
374	mov ebx, [fpu_owner]	374	mov ebx, [fpu_owner]
375	cmp ebx, [current_slot_idx]	375	cmp ebx, [current_slot_idx]
376	je .exit	376	je .exit
377		377
378	shl ebx, 8	378	shl ebx, BSF sizeof.APPDATA
379	mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state]	379	mov eax, [SLOT_BASE + ebx + APPDATA.fpu_state]
380	bt [cpu_caps+(CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32	380	bt [cpu_caps + (CAPS_OSXSAVE/32)*4], CAPS_OSXSAVE mod 32
381	jnc .no_xsave	381	jnc .no_xsave
382	mov ecx, eax	382	mov ecx, eax
383	mov eax, [xsave_eax]	383	mov eax, [xsave_eax]
384	mov edx, [xsave_edx]	384	mov edx, [xsave_edx]
385	xsave [ecx]	385	xsave [ecx]
386	mov ebx, [current_slot_idx]	386	mov ebx, [current_slot_idx]
387	mov [fpu_owner], ebx	387	mov [fpu_owner], ebx
388	shl ebx, 8	388	shl ebx, BSF sizeof.APPDATA
389	mov ecx, [ebx+SLOT_BASE+APPDATA.fpu_state]	389	mov ecx, [SLOT_BASE + ebx + APPDATA.fpu_state]
390	xrstor [ecx]	390	xrstor [ecx]
391	.exit:	391	.exit:
392	restore_ring3_context	392	restore_ring3_context
393	iret	393	iret
394	.no_xsave:	394	.no_xsave:
395	bt [cpu_caps], CAPS_SSE	395	bt [cpu_caps], CAPS_SSE
396	jnc .no_SSE	396	jnc .no_SSE
397		397
398	fxsave [eax]	398	fxsave [eax]
399	mov ebx, [current_slot_idx]	399	mov ebx, [current_slot_idx]
400	mov [fpu_owner], ebx	400	mov [fpu_owner], ebx
401	shl ebx, 8	401	shl ebx, BSF sizeof.APPDATA
402	mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state]	402	mov eax, [SLOT_BASE + ebx + APPDATA.fpu_state]
403	fxrstor [eax]	403	fxrstor [eax]
404	restore_ring3_context	404	restore_ring3_context
405	iret	405	iret
406		406
407	.no_SSE:	407	.no_SSE:
408	fnsave [eax]	408	fnsave [eax]
409	mov ebx, [current_slot_idx]	409	mov ebx, [current_slot_idx]
410	mov [fpu_owner], ebx	410	mov [fpu_owner], ebx
411	shl ebx, 8	411	shl ebx, BSF sizeof.APPDATA
412	mov eax, [ebx+SLOT_BASE+APPDATA.fpu_state]	412	mov eax, [SLOT_BASE + ebx + APPDATA.fpu_state]
413	frstor [eax]	413	frstor [eax]
414	restore_ring3_context	414	restore_ring3_context
415	iret	415	iret
416		416
417	iglobal	417	iglobal
418	fpu_owner dd 2	418	fpu_owner dd 2
419	endg	419	endg

Subversion Repositories Kolibri OS

(root)/kernel/trunk/core/fpu.inc @ 972 – Rev 8869 → 9715