1
dnl AMD K6 mpn_divrem_1 -- mpn by limb division.
3
dnl Copyright 1999, 2000, 2001, 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 2.1 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
18
dnl License along with the GNU MP Library; see the file COPYING.LIB. If
19
dnl not, write to the Free Software Foundation, Inc., 51 Franklin Street,
20
dnl Fifth Floor, Boston, MA 02110-1301, USA.
22
include(`../config.m4')
28
C mp_limb_t mpn_divrem_1 (mp_ptr dst, mp_size_t xsize,
29
C mp_srcptr src, mp_size_t size, mp_limb_t divisor);
30
C mp_limb_t mpn_divrem_1c (mp_ptr dst, mp_size_t xsize,
31
C mp_srcptr src, mp_size_t size, mp_limb_t divisor,
34
C The code here is basically the same as mpn/x86/divrem_1.asm, but uses loop
35
C instead of decl+jnz, since it comes out 2 cycles/limb faster.
37
C A test is done to see if the high limb is less the the divisor, and if so
38
C one less div is done. A div is 20 cycles, so assuming high<divisor about
39
C half the time, then this test saves half that amount. The branch
40
C misprediction penalty is less than that.
42
C Back-to-back div instructions run at 20 cycles, the same as the loop here,
43
C so it seems there's nothing to gain by rearranging the loop. Pairing the
44
C mov and loop instructions was found to gain nothing.
48
C The low-latency K6 multiply might be thought to suit a mul-by-inverse, but
49
C that algorithm has been found to suffer from the relatively poor carry
50
C handling on K6 and too many auxiliary instructions. The fractional part
51
C however could be done at about 13 c/l, if it mattered enough.
53
defframe(PARAM_CARRY, 24)
54
defframe(PARAM_DIVISOR,20)
55
defframe(PARAM_SIZE, 16)
56
defframe(PARAM_SRC, 12)
57
defframe(PARAM_XSIZE, 8)
58
defframe(PARAM_DST, 4)
63
PROLOGUE(mpn_divrem_1c)
67
pushl %edi FRAME_pushl()
70
pushl %esi FRAME_pushl()
72
movl PARAM_DIVISOR, %esi
73
pushl %ebx FRAME_pushl()
76
pushl %ebp FRAME_pushl()
78
movl PARAM_XSIZE, %ebp
81
movl PARAM_CARRY, %edx
82
jz L(fraction) C if size==0
84
leal -4(%ebx,%ebp,4), %ebx C dst one limb below integer part
91
PROLOGUE(mpn_divrem_1)
95
pushl %edi FRAME_pushl()
98
pushl %esi FRAME_pushl()
100
movl PARAM_DIVISOR, %esi
104
pushl %ebx FRAME_pushl()
106
movl -4(%edi,%ecx,4), %eax C src high limb
110
pushl %ebp FRAME_pushl()
112
movl PARAM_XSIZE, %ebp
115
leal -4(%ebx,%ebp,4), %ebx C dst one limb below integer part
119
C high<divisor, so high of dst is zero, and avoid one div
121
movl %edx, (%ebx,%ecx,4)
129
C eax scratch (quotient)
132
C edx scratch (remainder)
137
movl -4(%edi,%ecx,4), %eax
142
movl %eax, (%ebx,%ecx,4)
154
C eax scratch (quotient)
157
C edx scratch (remainder)
166
movl %eax, -4(%ebx,%ecx,4)
181
movl PARAM_XSIZE, %ecx
186
cld C better safe than sorry, see mpn/x86/README