1dnl  x86_64 mpn_bdiv_dbm1.
2
3dnl  Copyright 2008 Free Software Foundation, Inc.
4
5dnl  This file is part of the GNU MP Library.
6
7dnl  The GNU MP Library is free software; you can redistribute it and/or modify
8dnl  it under the terms of the GNU Lesser General Public License as published
9dnl  by the Free Software Foundation; either version 3 of the License, or (at
10dnl  your option) any later version.
11
12dnl  The GNU MP Library is distributed in the hope that it will be useful, but
13dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
14dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
15dnl  License for more details.
16
17dnl  You should have received a copy of the GNU Lesser General Public License
18dnl  along with the GNU MP Library.  If not, see http://www.gnu.org/licenses/.
19
20include(`../config.m4')
21
22C	     cycles/limb
23C K8,K9:	 2.25
24C K10:		  ?
25C P4:		12.5
26C P6 core2: 	 4.0
27C P6 corei7: 	 3.8
28C P6 atom:	20
29
30C TODO
31C  * Do proper 4-way feed-in instead of the current epilogue
32
33C INPUT PARAMETERS shared
34define(`qp',	`%rdi')
35define(`up',	`%rsi')
36define(`n',	`%rdx')
37define(`bd',	`%rcx')
38define(`cy',	`%r8')
39
40
41ASM_START()
42	TEXT
43	ALIGN(16)
44PROLOGUE(mpn_bdiv_dbm1c)
45	mov	(%rsi), %rax
46	mov	%rdx, %r9		C n
47
48	mul	%rcx
49	sub	%rax, %r8
50	mov	%r8, (%rdi)
51	sbb	%rdx, %r8
52
53	lea	(%rsi,%r9,8), %rsi
54	lea	(%rdi,%r9,8), %rdi
55	neg	%r9
56	add	$4, %r9
57	jns	L(end)
58	ALIGN(16)
59L(top):
60	mov	-24(%rsi,%r9,8), %rax
61	mul	%rcx
62	sub	%rax, %r8
63	mov	%r8, -24(%rdi,%r9,8)
64	sbb	%rdx, %r8
65L(3):
66	mov	-16(%rsi,%r9,8), %rax
67	mul	%rcx
68	sub	%rax, %r8
69	mov	%r8, -16(%rdi,%r9,8)
70	sbb	%rdx, %r8
71L(2):
72	mov	-8(%rsi,%r9,8), %rax
73	mul	%rcx
74	sub	%rax, %r8
75	mov	%r8, -8(%rdi,%r9,8)
76	sbb	%rdx, %r8
77L(1):
78	mov	(%rsi,%r9,8), %rax
79	mul	%rcx
80	sub	%rax, %r8
81	mov	%r8, (%rdi,%r9,8)
82	sbb	%rdx, %r8
83
84	add	$4, %r9
85	js	L(top)
86L(end):
87	je	L(3x)
88	cmp	$2, %r9
89	jg	L(ret)
90	mov	$-1, %r9
91	je	L(1)
92	jmp	L(2)
93L(3x):
94	dec	%r9
95	jmp	L(3)
96
97L(ret):	mov	%r8, %rax
98	ret
99EPILOGUE()
100