1*5ba6b03cSchristos; inffasx64.asm is a hand tuned assembler version of inffast.c - fast decoding
2*5ba6b03cSchristos; version for AMD64 on Windows using Microsoft C compiler
3*5ba6b03cSchristos;
4*5ba6b03cSchristos; inffasx64.asm is automatically convert from AMD64 portion of inffas86.c
5*5ba6b03cSchristos; inffasx64.asm is called by inffas8664.c, which contain more info.
6*5ba6b03cSchristos
7*5ba6b03cSchristos
8*5ba6b03cSchristos; to compile this file, I use option
9*5ba6b03cSchristos;   ml64.exe /Flinffasx64 /c /Zi inffasx64.asm
10*5ba6b03cSchristos;   with Microsoft Macro Assembler (x64) for AMD64
11*5ba6b03cSchristos;
12*5ba6b03cSchristos
13*5ba6b03cSchristos; This file compile with Microsoft Macro Assembler (x64) for AMD64
14*5ba6b03cSchristos;
15*5ba6b03cSchristos;   ml64.exe is given with Visual Studio 2005/2008/2010 and Windows WDK
16*5ba6b03cSchristos;
17*5ba6b03cSchristos;   (you can get Windows WDK with ml64 for AMD64 from
18*5ba6b03cSchristos;      http://www.microsoft.com/whdc/Devtools/wdk/default.mspx for low price)
19*5ba6b03cSchristos;
20*5ba6b03cSchristos
21*5ba6b03cSchristos
22*5ba6b03cSchristos.code
23*5ba6b03cSchristosinffas8664fnc PROC
24*5ba6b03cSchristos
25*5ba6b03cSchristos; see http://weblogs.asp.net/oldnewthing/archive/2004/01/14/58579.aspx and
26*5ba6b03cSchristos; http://msdn.microsoft.com/library/en-us/kmarch/hh/kmarch/64bitAMD_8e951dd2-ee77-4728-8702-55ce4b5dd24a.xml.asp
27*5ba6b03cSchristos;
28*5ba6b03cSchristos; All registers must be preserved across the call, except for
29*5ba6b03cSchristos;   rax, rcx, rdx, r8, r-9, r10, and r11, which are scratch.
30*5ba6b03cSchristos
31*5ba6b03cSchristos
32*5ba6b03cSchristos	mov [rsp-8],rsi
33*5ba6b03cSchristos	mov [rsp-16],rdi
34*5ba6b03cSchristos	mov [rsp-24],r12
35*5ba6b03cSchristos	mov [rsp-32],r13
36*5ba6b03cSchristos	mov [rsp-40],r14
37*5ba6b03cSchristos	mov [rsp-48],r15
38*5ba6b03cSchristos	mov [rsp-56],rbx
39*5ba6b03cSchristos
40*5ba6b03cSchristos	mov rax,rcx
41*5ba6b03cSchristos
42*5ba6b03cSchristos	mov	[rax+8], rbp       ; /* save regs rbp and rsp */
43*5ba6b03cSchristos	mov	[rax], rsp
44*5ba6b03cSchristos
45*5ba6b03cSchristos	mov	rsp, rax          ; /* make rsp point to &ar */
46*5ba6b03cSchristos
47*5ba6b03cSchristos	mov	rsi, [rsp+16]      ; /* rsi  = in */
48*5ba6b03cSchristos	mov	rdi, [rsp+32]      ; /* rdi  = out */
49*5ba6b03cSchristos	mov	r9, [rsp+24]       ; /* r9   = last */
50*5ba6b03cSchristos	mov	r10, [rsp+48]      ; /* r10  = end */
51*5ba6b03cSchristos	mov	rbp, [rsp+64]      ; /* rbp  = lcode */
52*5ba6b03cSchristos	mov	r11, [rsp+72]      ; /* r11  = dcode */
53*5ba6b03cSchristos	mov	rdx, [rsp+80]      ; /* rdx  = hold */
54*5ba6b03cSchristos	mov	ebx, [rsp+88]      ; /* ebx  = bits */
55*5ba6b03cSchristos	mov	r12d, [rsp+100]    ; /* r12d = lmask */
56*5ba6b03cSchristos	mov	r13d, [rsp+104]    ; /* r13d = dmask */
57*5ba6b03cSchristos                                          ; /* r14d = len */
58*5ba6b03cSchristos                                          ; /* r15d = dist */
59*5ba6b03cSchristos
60*5ba6b03cSchristos
61*5ba6b03cSchristos	cld
62*5ba6b03cSchristos	cmp	r10, rdi
63*5ba6b03cSchristos	je	L_one_time           ; /* if only one decode left */
64*5ba6b03cSchristos	cmp	r9, rsi
65*5ba6b03cSchristos
66*5ba6b03cSchristos    jne L_do_loop
67*5ba6b03cSchristos
68*5ba6b03cSchristos
69*5ba6b03cSchristosL_one_time:
70*5ba6b03cSchristos	mov	r8, r12           ; /* r8 = lmask */
71*5ba6b03cSchristos	cmp	bl, 32
72*5ba6b03cSchristos	ja	L_get_length_code_one_time
73*5ba6b03cSchristos
74*5ba6b03cSchristos	lodsd                         ; /* eax = *(uint *)in++ */
75*5ba6b03cSchristos	mov	cl, bl            ; /* cl = bits, needs it for shifting */
76*5ba6b03cSchristos	add	bl, 32             ; /* bits += 32 */
77*5ba6b03cSchristos	shl	rax, cl
78*5ba6b03cSchristos	or	rdx, rax          ; /* hold |= *((uint *)in)++ << bits */
79*5ba6b03cSchristos	jmp	L_get_length_code_one_time
80*5ba6b03cSchristos
81*5ba6b03cSchristosALIGN 4
82*5ba6b03cSchristosL_while_test:
83*5ba6b03cSchristos	cmp	r10, rdi
84*5ba6b03cSchristos	jbe	L_break_loop
85*5ba6b03cSchristos	cmp	r9, rsi
86*5ba6b03cSchristos	jbe	L_break_loop
87*5ba6b03cSchristos
88*5ba6b03cSchristosL_do_loop:
89*5ba6b03cSchristos	mov	r8, r12           ; /* r8 = lmask */
90*5ba6b03cSchristos	cmp	bl, 32
91*5ba6b03cSchristos	ja	L_get_length_code    ; /* if (32 < bits) */
92*5ba6b03cSchristos
93*5ba6b03cSchristos	lodsd                         ; /* eax = *(uint *)in++ */
94*5ba6b03cSchristos	mov	cl, bl            ; /* cl = bits, needs it for shifting */
95*5ba6b03cSchristos	add	bl, 32             ; /* bits += 32 */
96*5ba6b03cSchristos	shl	rax, cl
97*5ba6b03cSchristos	or	rdx, rax          ; /* hold |= *((uint *)in)++ << bits */
98*5ba6b03cSchristos
99*5ba6b03cSchristosL_get_length_code:
100*5ba6b03cSchristos	and	r8, rdx            ; /* r8 &= hold */
101*5ba6b03cSchristos	mov	eax, [rbp+r8*4]  ; /* eax = lcode[hold & lmask] */
102*5ba6b03cSchristos
103*5ba6b03cSchristos	mov	cl, ah            ; /* cl = this.bits */
104*5ba6b03cSchristos	sub	bl, ah            ; /* bits -= this.bits */
105*5ba6b03cSchristos	shr	rdx, cl           ; /* hold >>= this.bits */
106*5ba6b03cSchristos
107*5ba6b03cSchristos	test	al, al
108*5ba6b03cSchristos	jnz	L_test_for_length_base ; /* if (op != 0) 45.7% */
109*5ba6b03cSchristos
110*5ba6b03cSchristos	mov	r8, r12            ; /* r8 = lmask */
111*5ba6b03cSchristos	shr	eax, 16            ; /* output this.val char */
112*5ba6b03cSchristos	stosb
113*5ba6b03cSchristos
114*5ba6b03cSchristosL_get_length_code_one_time:
115*5ba6b03cSchristos	and	r8, rdx            ; /* r8 &= hold */
116*5ba6b03cSchristos	mov	eax, [rbp+r8*4] ; /* eax = lcode[hold & lmask] */
117*5ba6b03cSchristos
118*5ba6b03cSchristosL_dolen:
119*5ba6b03cSchristos	mov	cl, ah            ; /* cl = this.bits */
120*5ba6b03cSchristos	sub	bl, ah            ; /* bits -= this.bits */
121*5ba6b03cSchristos	shr	rdx, cl           ; /* hold >>= this.bits */
122*5ba6b03cSchristos
123*5ba6b03cSchristos	test	al, al
124*5ba6b03cSchristos	jnz	L_test_for_length_base ; /* if (op != 0) 45.7% */
125*5ba6b03cSchristos
126*5ba6b03cSchristos	shr	eax, 16            ; /* output this.val char */
127*5ba6b03cSchristos	stosb
128*5ba6b03cSchristos	jmp	L_while_test
129*5ba6b03cSchristos
130*5ba6b03cSchristosALIGN 4
131*5ba6b03cSchristosL_test_for_length_base:
132*5ba6b03cSchristos	mov	r14d, eax         ; /* len = this */
133*5ba6b03cSchristos	shr	r14d, 16           ; /* len = this.val */
134*5ba6b03cSchristos	mov	cl, al
135*5ba6b03cSchristos
136*5ba6b03cSchristos	test	al, 16
137*5ba6b03cSchristos	jz	L_test_for_second_level_length ; /* if ((op & 16) == 0) 8% */
138*5ba6b03cSchristos	and	cl, 15             ; /* op &= 15 */
139*5ba6b03cSchristos	jz	L_decode_distance    ; /* if (!op) */
140*5ba6b03cSchristos
141*5ba6b03cSchristosL_add_bits_to_len:
142*5ba6b03cSchristos	sub	bl, cl
143*5ba6b03cSchristos	xor	eax, eax
144*5ba6b03cSchristos	inc	eax
145*5ba6b03cSchristos	shl	eax, cl
146*5ba6b03cSchristos	dec	eax
147*5ba6b03cSchristos	and	eax, edx          ; /* eax &= hold */
148*5ba6b03cSchristos	shr	rdx, cl
149*5ba6b03cSchristos	add	r14d, eax         ; /* len += hold & mask[op] */
150*5ba6b03cSchristos
151*5ba6b03cSchristosL_decode_distance:
152*5ba6b03cSchristos	mov	r8, r13           ; /* r8 = dmask */
153*5ba6b03cSchristos	cmp	bl, 32
154*5ba6b03cSchristos	ja	L_get_distance_code  ; /* if (32 < bits) */
155*5ba6b03cSchristos
156*5ba6b03cSchristos	lodsd                         ; /* eax = *(uint *)in++ */
157*5ba6b03cSchristos	mov	cl, bl            ; /* cl = bits, needs it for shifting */
158*5ba6b03cSchristos	add	bl, 32             ; /* bits += 32 */
159*5ba6b03cSchristos	shl	rax, cl
160*5ba6b03cSchristos	or	rdx, rax          ; /* hold |= *((uint *)in)++ << bits */
161*5ba6b03cSchristos
162*5ba6b03cSchristosL_get_distance_code:
163*5ba6b03cSchristos	and	r8, rdx           ; /* r8 &= hold */
164*5ba6b03cSchristos	mov	eax, [r11+r8*4] ; /* eax = dcode[hold & dmask] */
165*5ba6b03cSchristos
166*5ba6b03cSchristosL_dodist:
167*5ba6b03cSchristos	mov	r15d, eax         ; /* dist = this */
168*5ba6b03cSchristos	shr	r15d, 16           ; /* dist = this.val */
169*5ba6b03cSchristos	mov	cl, ah
170*5ba6b03cSchristos	sub	bl, ah            ; /* bits -= this.bits */
171*5ba6b03cSchristos	shr	rdx, cl           ; /* hold >>= this.bits */
172*5ba6b03cSchristos	mov	cl, al            ; /* cl = this.op */
173*5ba6b03cSchristos
174*5ba6b03cSchristos	test	al, 16             ; /* if ((op & 16) == 0) */
175*5ba6b03cSchristos	jz	L_test_for_second_level_dist
176*5ba6b03cSchristos	and	cl, 15             ; /* op &= 15 */
177*5ba6b03cSchristos	jz	L_check_dist_one
178*5ba6b03cSchristos
179*5ba6b03cSchristosL_add_bits_to_dist:
180*5ba6b03cSchristos	sub	bl, cl
181*5ba6b03cSchristos	xor	eax, eax
182*5ba6b03cSchristos	inc	eax
183*5ba6b03cSchristos	shl	eax, cl
184*5ba6b03cSchristos	dec	eax                 ; /* (1 << op) - 1 */
185*5ba6b03cSchristos	and	eax, edx          ; /* eax &= hold */
186*5ba6b03cSchristos	shr	rdx, cl
187*5ba6b03cSchristos	add	r15d, eax         ; /* dist += hold & ((1 << op) - 1) */
188*5ba6b03cSchristos
189*5ba6b03cSchristosL_check_window:
190*5ba6b03cSchristos	mov	r8, rsi           ; /* save in so from can use it's reg */
191*5ba6b03cSchristos	mov	rax, rdi
192*5ba6b03cSchristos	sub	rax, [rsp+40]      ; /* nbytes = out - beg */
193*5ba6b03cSchristos
194*5ba6b03cSchristos	cmp	eax, r15d
195*5ba6b03cSchristos	jb	L_clip_window        ; /* if (dist > nbytes) 4.2% */
196*5ba6b03cSchristos
197*5ba6b03cSchristos	mov	ecx, r14d         ; /* ecx = len */
198*5ba6b03cSchristos	mov	rsi, rdi
199*5ba6b03cSchristos	sub	rsi, r15          ; /* from = out - dist */
200*5ba6b03cSchristos
201*5ba6b03cSchristos	sar	ecx, 1
202*5ba6b03cSchristos	jnc	L_copy_two           ; /* if len % 2 == 0 */
203*5ba6b03cSchristos
204*5ba6b03cSchristos	rep     movsw
205*5ba6b03cSchristos	mov	al, [rsi]
206*5ba6b03cSchristos	mov	[rdi], al
207*5ba6b03cSchristos	inc	rdi
208*5ba6b03cSchristos
209*5ba6b03cSchristos	mov	rsi, r8           ; /* move in back to %rsi, toss from */
210*5ba6b03cSchristos	jmp	L_while_test
211*5ba6b03cSchristos
212*5ba6b03cSchristosL_copy_two:
213*5ba6b03cSchristos	rep     movsw
214*5ba6b03cSchristos	mov	rsi, r8           ; /* move in back to %rsi, toss from */
215*5ba6b03cSchristos	jmp	L_while_test
216*5ba6b03cSchristos
217*5ba6b03cSchristosALIGN 4
218*5ba6b03cSchristosL_check_dist_one:
219*5ba6b03cSchristos	cmp	r15d, 1            ; /* if dist 1, is a memset */
220*5ba6b03cSchristos	jne	L_check_window
221*5ba6b03cSchristos	cmp	[rsp+40], rdi      ; /* if out == beg, outside window */
222*5ba6b03cSchristos	je	L_check_window
223*5ba6b03cSchristos
224*5ba6b03cSchristos	mov	ecx, r14d         ; /* ecx = len */
225*5ba6b03cSchristos	mov	al, [rdi-1]
226*5ba6b03cSchristos	mov	ah, al
227*5ba6b03cSchristos
228*5ba6b03cSchristos	sar	ecx, 1
229*5ba6b03cSchristos	jnc	L_set_two
230*5ba6b03cSchristos	mov	[rdi], al
231*5ba6b03cSchristos	inc	rdi
232*5ba6b03cSchristos
233*5ba6b03cSchristosL_set_two:
234*5ba6b03cSchristos	rep     stosw
235*5ba6b03cSchristos	jmp	L_while_test
236*5ba6b03cSchristos
237*5ba6b03cSchristosALIGN 4
238*5ba6b03cSchristosL_test_for_second_level_length:
239*5ba6b03cSchristos	test	al, 64
240*5ba6b03cSchristos	jnz	L_test_for_end_of_block ; /* if ((op & 64) != 0) */
241*5ba6b03cSchristos
242*5ba6b03cSchristos	xor	eax, eax
243*5ba6b03cSchristos	inc	eax
244*5ba6b03cSchristos	shl	eax, cl
245*5ba6b03cSchristos	dec	eax
246*5ba6b03cSchristos	and	eax, edx         ; /* eax &= hold */
247*5ba6b03cSchristos	add	eax, r14d        ; /* eax += len */
248*5ba6b03cSchristos	mov	eax, [rbp+rax*4] ; /* eax = lcode[val+(hold&mask[op])]*/
249*5ba6b03cSchristos	jmp	L_dolen
250*5ba6b03cSchristos
251*5ba6b03cSchristosALIGN 4
252*5ba6b03cSchristosL_test_for_second_level_dist:
253*5ba6b03cSchristos	test	al, 64
254*5ba6b03cSchristos	jnz	L_invalid_distance_code ; /* if ((op & 64) != 0) */
255*5ba6b03cSchristos
256*5ba6b03cSchristos	xor	eax, eax
257*5ba6b03cSchristos	inc	eax
258*5ba6b03cSchristos	shl	eax, cl
259*5ba6b03cSchristos	dec	eax
260*5ba6b03cSchristos	and	eax, edx         ; /* eax &= hold */
261*5ba6b03cSchristos	add	eax, r15d        ; /* eax += dist */
262*5ba6b03cSchristos	mov	eax, [r11+rax*4] ; /* eax = dcode[val+(hold&mask[op])]*/
263*5ba6b03cSchristos	jmp	L_dodist
264*5ba6b03cSchristos
265*5ba6b03cSchristosALIGN 4
266*5ba6b03cSchristosL_clip_window:
267*5ba6b03cSchristos	mov	ecx, eax         ; /* ecx = nbytes */
268*5ba6b03cSchristos	mov	eax, [rsp+92]     ; /* eax = wsize, prepare for dist cmp */
269*5ba6b03cSchristos	neg	ecx                ; /* nbytes = -nbytes */
270*5ba6b03cSchristos
271*5ba6b03cSchristos	cmp	eax, r15d
272*5ba6b03cSchristos	jb	L_invalid_distance_too_far ; /* if (dist > wsize) */
273*5ba6b03cSchristos
274*5ba6b03cSchristos	add	ecx, r15d         ; /* nbytes = dist - nbytes */
275*5ba6b03cSchristos	cmp	dword ptr [rsp+96], 0
276*5ba6b03cSchristos	jne	L_wrap_around_window ; /* if (write != 0) */
277*5ba6b03cSchristos
278*5ba6b03cSchristos	mov	rsi, [rsp+56]     ; /* from  = window */
279*5ba6b03cSchristos	sub	eax, ecx         ; /* eax  -= nbytes */
280*5ba6b03cSchristos	add	rsi, rax         ; /* from += wsize - nbytes */
281*5ba6b03cSchristos
282*5ba6b03cSchristos	mov	eax, r14d        ; /* eax = len */
283*5ba6b03cSchristos	cmp	r14d, ecx
284*5ba6b03cSchristos	jbe	L_do_copy           ; /* if (nbytes >= len) */
285*5ba6b03cSchristos
286*5ba6b03cSchristos	sub	eax, ecx         ; /* eax -= nbytes */
287*5ba6b03cSchristos	rep     movsb
288*5ba6b03cSchristos	mov	rsi, rdi
289*5ba6b03cSchristos	sub	rsi, r15         ; /* from = &out[ -dist ] */
290*5ba6b03cSchristos	jmp	L_do_copy
291*5ba6b03cSchristos
292*5ba6b03cSchristosALIGN 4
293*5ba6b03cSchristosL_wrap_around_window:
294*5ba6b03cSchristos	mov	eax, [rsp+96]     ; /* eax = write */
295*5ba6b03cSchristos	cmp	ecx, eax
296*5ba6b03cSchristos	jbe	L_contiguous_in_window ; /* if (write >= nbytes) */
297*5ba6b03cSchristos
298*5ba6b03cSchristos	mov	esi, [rsp+92]     ; /* from  = wsize */
299*5ba6b03cSchristos	add	rsi, [rsp+56]     ; /* from += window */
300*5ba6b03cSchristos	add	rsi, rax         ; /* from += write */
301*5ba6b03cSchristos	sub	rsi, rcx         ; /* from -= nbytes */
302*5ba6b03cSchristos	sub	ecx, eax         ; /* nbytes -= write */
303*5ba6b03cSchristos
304*5ba6b03cSchristos	mov	eax, r14d        ; /* eax = len */
305*5ba6b03cSchristos	cmp	eax, ecx
306*5ba6b03cSchristos	jbe	L_do_copy           ; /* if (nbytes >= len) */
307*5ba6b03cSchristos
308*5ba6b03cSchristos	sub	eax, ecx         ; /* len -= nbytes */
309*5ba6b03cSchristos	rep     movsb
310*5ba6b03cSchristos	mov	rsi, [rsp+56]     ; /* from = window */
311*5ba6b03cSchristos	mov	ecx, [rsp+96]     ; /* nbytes = write */
312*5ba6b03cSchristos	cmp	eax, ecx
313*5ba6b03cSchristos	jbe	L_do_copy           ; /* if (nbytes >= len) */
314*5ba6b03cSchristos
315*5ba6b03cSchristos	sub	eax, ecx         ; /* len -= nbytes */
316*5ba6b03cSchristos	rep     movsb
317*5ba6b03cSchristos	mov	rsi, rdi
318*5ba6b03cSchristos	sub	rsi, r15         ; /* from = out - dist */
319*5ba6b03cSchristos	jmp	L_do_copy
320*5ba6b03cSchristos
321*5ba6b03cSchristosALIGN 4
322*5ba6b03cSchristosL_contiguous_in_window:
323*5ba6b03cSchristos	mov	rsi, [rsp+56]     ; /* rsi = window */
324*5ba6b03cSchristos	add	rsi, rax
325*5ba6b03cSchristos	sub	rsi, rcx         ; /* from += write - nbytes */
326*5ba6b03cSchristos
327*5ba6b03cSchristos	mov	eax, r14d        ; /* eax = len */
328*5ba6b03cSchristos	cmp	eax, ecx
329*5ba6b03cSchristos	jbe	L_do_copy           ; /* if (nbytes >= len) */
330*5ba6b03cSchristos
331*5ba6b03cSchristos	sub	eax, ecx         ; /* len -= nbytes */
332*5ba6b03cSchristos	rep     movsb
333*5ba6b03cSchristos	mov	rsi, rdi
334*5ba6b03cSchristos	sub	rsi, r15         ; /* from = out - dist */
335*5ba6b03cSchristos	jmp	L_do_copy           ; /* if (nbytes >= len) */
336*5ba6b03cSchristos
337*5ba6b03cSchristosALIGN 4
338*5ba6b03cSchristosL_do_copy:
339*5ba6b03cSchristos	mov	ecx, eax         ; /* ecx = len */
340*5ba6b03cSchristos	rep     movsb
341*5ba6b03cSchristos
342*5ba6b03cSchristos	mov	rsi, r8          ; /* move in back to %esi, toss from */
343*5ba6b03cSchristos	jmp	L_while_test
344*5ba6b03cSchristos
345*5ba6b03cSchristosL_test_for_end_of_block:
346*5ba6b03cSchristos	test	al, 32
347*5ba6b03cSchristos	jz	L_invalid_literal_length_code
348*5ba6b03cSchristos	mov	dword ptr [rsp+116], 1
349*5ba6b03cSchristos	jmp	L_break_loop_with_status
350*5ba6b03cSchristos
351*5ba6b03cSchristosL_invalid_literal_length_code:
352*5ba6b03cSchristos	mov	dword ptr [rsp+116], 2
353*5ba6b03cSchristos	jmp	L_break_loop_with_status
354*5ba6b03cSchristos
355*5ba6b03cSchristosL_invalid_distance_code:
356*5ba6b03cSchristos	mov	dword ptr [rsp+116], 3
357*5ba6b03cSchristos	jmp	L_break_loop_with_status
358*5ba6b03cSchristos
359*5ba6b03cSchristosL_invalid_distance_too_far:
360*5ba6b03cSchristos	mov	dword ptr [rsp+116], 4
361*5ba6b03cSchristos	jmp	L_break_loop_with_status
362*5ba6b03cSchristos
363*5ba6b03cSchristosL_break_loop:
364*5ba6b03cSchristos	mov	dword ptr [rsp+116], 0
365*5ba6b03cSchristos
366*5ba6b03cSchristosL_break_loop_with_status:
367*5ba6b03cSchristos; /* put in, out, bits, and hold back into ar and pop esp */
368*5ba6b03cSchristos	mov	[rsp+16], rsi     ; /* in */
369*5ba6b03cSchristos	mov	[rsp+32], rdi     ; /* out */
370*5ba6b03cSchristos	mov	[rsp+88], ebx     ; /* bits */
371*5ba6b03cSchristos	mov	[rsp+80], rdx     ; /* hold */
372*5ba6b03cSchristos
373*5ba6b03cSchristos	mov	rax, [rsp]       ; /* restore rbp and rsp */
374*5ba6b03cSchristos	mov	rbp, [rsp+8]
375*5ba6b03cSchristos	mov	rsp, rax
376*5ba6b03cSchristos
377*5ba6b03cSchristos
378*5ba6b03cSchristos
379*5ba6b03cSchristos	mov rsi,[rsp-8]
380*5ba6b03cSchristos	mov rdi,[rsp-16]
381*5ba6b03cSchristos	mov r12,[rsp-24]
382*5ba6b03cSchristos	mov r13,[rsp-32]
383*5ba6b03cSchristos	mov r14,[rsp-40]
384*5ba6b03cSchristos	mov r15,[rsp-48]
385*5ba6b03cSchristos	mov rbx,[rsp-56]
386*5ba6b03cSchristos
387*5ba6b03cSchristos    ret 0
388*5ba6b03cSchristos;          :
389*5ba6b03cSchristos;          : "m" (ar)
390*5ba6b03cSchristos;          : "memory", "%rax", "%rbx", "%rcx", "%rdx", "%rsi", "%rdi",
391*5ba6b03cSchristos;            "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15"
392*5ba6b03cSchristos;    );
393*5ba6b03cSchristos
394*5ba6b03cSchristosinffas8664fnc 	ENDP
395*5ba6b03cSchristos;_TEXT	ENDS
396*5ba6b03cSchristosEND
397