Annotation of OpenXM_contrib/gmp/mpn/x86/k7/mmx/copyd.asm, Revision 1.1.1.1
1.1 maekawa 1: dnl AMD K7 mpn_copyd -- copy limb vector, decrementing.
2: dnl
3: dnl alignment dst/src, A=0mod8 N=4mod8
4: dnl A/A A/N N/A N/N
5: dnl K7 0.75 1.0 1.0 0.75
6:
7:
8: dnl Copyright (C) 1999, 2000 Free Software Foundation, Inc.
9: dnl
10: dnl This file is part of the GNU MP Library.
11: dnl
12: dnl The GNU MP Library is free software; you can redistribute it and/or
13: dnl modify it under the terms of the GNU Lesser General Public License as
14: dnl published by the Free Software Foundation; either version 2.1 of the
15: dnl License, or (at your option) any later version.
16: dnl
17: dnl The GNU MP Library is distributed in the hope that it will be useful,
18: dnl but WITHOUT ANY WARRANTY; without even the implied warranty of
19: dnl MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20: dnl Lesser General Public License for more details.
21: dnl
22: dnl You should have received a copy of the GNU Lesser General Public
23: dnl License along with the GNU MP Library; see the file COPYING.LIB. If
24: dnl not, write to the Free Software Foundation, Inc., 59 Temple Place -
25: dnl Suite 330, Boston, MA 02111-1307, USA.
26:
27:
28: include(`../config.m4')
29:
30:
31: C void mpn_copyd (mp_ptr dst, mp_srcptr src, mp_size_t size);
32: C
33: C The various comments in mpn/x86/k7/copyi.asm apply here too.
34:
35: defframe(PARAM_SIZE,12)
36: defframe(PARAM_SRC, 8)
37: defframe(PARAM_DST, 4)
38: deflit(`FRAME',0)
39:
40: dnl parameter space reused
41: define(SAVE_EBX,`PARAM_SIZE')
42: define(SAVE_ESI,`PARAM_SRC')
43:
44: dnl minimum 5 since the unrolled code can't handle less than 5
45: deflit(UNROLL_THRESHOLD, 5)
46:
47: .text
48: ALIGN(32)
49: PROLOGUE(mpn_copyd)
50:
51: movl PARAM_SIZE, %ecx
52: movl %ebx, SAVE_EBX
53:
54: movl PARAM_SRC, %eax
55: movl PARAM_DST, %edx
56:
57: cmpl $UNROLL_THRESHOLD, %ecx
58: jae L(unroll)
59:
60: orl %ecx, %ecx
61: jz L(simple_done)
62:
63: L(simple):
64: C eax src
65: C ebx scratch
66: C ecx counter
67: C edx dst
68: C
69: C this loop is 2 cycles/limb
70:
71: movl -4(%eax,%ecx,4), %ebx
72: movl %ebx, -4(%edx,%ecx,4)
73: decl %ecx
74: jnz L(simple)
75:
76: L(simple_done):
77: movl SAVE_EBX, %ebx
78: ret
79:
80:
81: L(unroll):
82: movl %esi, SAVE_ESI
83: leal (%eax,%ecx,4), %ebx
84: leal (%edx,%ecx,4), %esi
85:
86: andl %esi, %ebx
87: movl SAVE_ESI, %esi
88: subl $4, %ecx C size-4
89:
90: testl $4, %ebx C testl to pad code closer to 16 bytes for L(top)
91: jz L(aligned)
92:
93: C both src and dst unaligned, process one limb to align them
94: movl 12(%eax,%ecx,4), %ebx
95: movl %ebx, 12(%edx,%ecx,4)
96: decl %ecx
97: L(aligned):
98:
99:
100: ALIGN(16)
101: L(top):
102: C eax src
103: C ebx
104: C ecx counter, limbs
105: C edx dst
106:
107: movq 8(%eax,%ecx,4), %mm0
108: movq (%eax,%ecx,4), %mm1
109: subl $4, %ecx
110: movq %mm0, 16+8(%edx,%ecx,4)
111: movq %mm1, 16(%edx,%ecx,4)
112: jns L(top)
113:
114:
115: C now %ecx is -4 to -1 representing respectively 0 to 3 limbs remaining
116:
117: testb $2, %cl
118: jz L(finish_not_two)
119:
120: movq 8(%eax,%ecx,4), %mm0
121: movq %mm0, 8(%edx,%ecx,4)
122: L(finish_not_two):
123:
124: testb $1, %cl
125: jz L(done)
126:
127: movl (%eax), %ebx
128: movl %ebx, (%edx)
129:
130: L(done):
131: movl SAVE_EBX, %ebx
132: emms
133: ret
134:
135:
136: EPILOGUE()
FreeBSD-CVSweb <freebsd-cvsweb@FreeBSD.org>