1 dnl AMD K7 mpn_copyi -- copy limb vector, incrementing.
3 dnl Copyright 1999, 2000, 2002, 2003 Free Software Foundation, Inc.
5 dnl This file is part of the GNU MP Library.
7 dnl The GNU MP Library is free software; you can redistribute it and/or
8 dnl modify it under the terms of the GNU Lesser General Public License as
9 dnl published by the Free Software Foundation; either version 3 of the
10 dnl License, or (at your option) any later version.
12 dnl The GNU MP Library is distributed in the hope that it will be useful,
13 dnl but WITHOUT ANY WARRANTY; without even the implied warranty of
14 dnl MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 dnl Lesser General Public License for more details.
17 dnl You should have received a copy of the GNU Lesser General Public License
18 dnl along with the GNU MP Library. If not, see http://www.gnu.org/licenses/.
20 include(`../config.m4')
23 C alignment dst/src, A=0mod8 N=4mod8
25 C K7 0.75 1.0 1.0 0.75
28 C void mpn_copyi (mp_ptr dst, mp_srcptr src, mp_size_t size);
30 C Copy src,size to dst,size.
32 C This code at 0.75 or 1.0 c/l is always faster than a plain rep movsl at
35 C The K7 can do a 64-bit load and 64-bit store in one cycle (optimization
36 C guile 22007 appendix B), so 0.5 c/l should be possible, however nothing
37 C under 0.7 c/l is known. Apparently only two 32-bit stores can be done in
38 C one cycle, so perhaps some scheduling is needed to ensure it's a
39 C load+store in each cycle, not store+store.
41 C If both source and destination are unaligned then one limb is processed at
42 C the start to make them aligned and so get 0.75 c/l, whereas if they'd been
43 C used unaligned it would be 1.5 c/l.
45 defframe(PARAM_SIZE,12)
46 defframe(PARAM_SRC, 8)
47 defframe(PARAM_DST, 4)
49 dnl parameter space reused
50 define(SAVE_EBX,`PARAM_SIZE')
52 dnl minimum 5 since the unrolled code can't handle less than 5
53 deflit(UNROLL_THRESHOLD, 5)
66 cmpl $UNROLL_THRESHOLD, %ecx
73 C eax src, incrementing
76 C edx dst, incrementing
78 C this loop is 2 cycles/limb
94 leal -12(%eax,%ecx,4), %eax C src end - 12
95 subl $3, %ecx C size-3
98 leal (%edx,%ecx,4), %edx C dst end - 12
101 testl $4, %ebx C testl to pad code closer to 16 bytes for L(top)
104 C both src and dst unaligned, process one limb to align them
105 movl (%eax,%ecx,4), %ebx
106 movl %ebx, (%edx,%ecx,4)
115 C ecx counter, negative, limbs
118 movq (%eax,%ecx,4), %mm0
119 movq 8(%eax,%ecx,4), %mm1
121 movq %mm0, -16(%edx,%ecx,4)
122 movq %mm1, -16+8(%edx,%ecx,4)
123 ja L(top) C jump no carry and not zero
126 C now %ecx is 0 to 3 representing respectively 3 to 0 limbs remaining
129 jnz L(finish_not_two)
131 movq (%eax,%ecx,4), %mm0
132 movq %mm0, (%edx,%ecx,4)