1 dnl x86 mpn_divrem_1
-- mpn by limb division extending to fractional quotient.
3 dnl Copyright
1999-2003, 2007 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 modify
8 dnl it under the terms of
either:
10 dnl
* the GNU Lesser General
Public License as published by the Free
11 dnl Software Foundation
; either version 3 of the License, or (at your
12 dnl option
) any later version.
16 dnl
* the GNU General
Public License as published by the Free Software
17 dnl Foundation
; either version 2 of the License, or (at your option) any
20 dnl
or both
in parallel
, as here.
22 dnl The GNU MP Library is distributed
in the hope that it will be useful
, but
23 dnl WITHOUT ANY WARRANTY
; without even the implied warranty of MERCHANTABILITY
24 dnl
or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General
Public License
27 dnl You should have received copies of the GNU General
Public License
and the
28 dnl GNU Lesser General
Public License along with the GNU MP Library. If
not,
29 dnl see
https://www.gnu.
org/licenses
/.
31 include(`..
/config.m4
')
44 C mp_limb_t mpn_divrem_1 (mp_ptr dst, mp_size_t xsize,
45 C mp_srcptr src, mp_size_t size, mp_limb_t divisor);
46 C mp_limb_t mpn_divrem_1c (mp_ptr dst, mp_size_t xsize,
47 C mp_srcptr src, mp_size_t size, mp_limb_t divisor,
50 C Divide src,size by divisor and store the quotient in dst+xsize,size.
51 C Extend the division to fractional quotient limbs in dst,xsize. Return the
52 C remainder. Either or both xsize and size can be 0.
54 C mpn_divrem_1c takes a carry parameter which is an initial high limb,
55 C effectively one extra limb at the top of src,size. Must have
59 C Essentially the code is the same as the division based part of
60 C mpn/generic/divrem_1.c, but has the advantage that we get the desired divl
61 C instruction even when gcc is not being used (when longlong.h only has the
62 C rather slow generic C udiv_qrnnd().
64 C A test is done to see if the high limb is less than the divisor, and if so
65 C one less div is done. A div is between 20 and 40 cycles on the various
66 C x86s, so assuming high<divisor about half the time, then this test saves
67 C half that amount. The branch misprediction penalty on each chip is less
73 C It might be thought that moving the load down to pair with the store would
74 C save 1 cycle, but that doesn't seem to happen
in practice
, and in any case
75 C would be a mere
2.2% saving
, so it
's hardly worth bothering about.
77 C A mul-by-inverse might be a possibility for P5, as done in
78 C mpn/x86/pentium/mod_1.asm. The number of auxiliary instructions required
79 C is a hinderance, but there could be a 10-15% speedup available.
84 C K6 has its own version of this code, using loop and paying attention to
85 C cache line boundary crossings. The target 20 c/l can be had with the
86 C decl+jnz of the present code by pairing up the load and store in the
87 C loops. But it's considered easier
not to introduce complexity just for
88 C that
, but instead let k6 have its own code.
91 defframe
(PARAM_CARRY
, 24)
92 defframe
(PARAM_DIVISOR
,20)
93 defframe
(PARAM_SIZE
, 16)
94 defframe
(PARAM_SRC
, 12)
95 defframe
(PARAM_XSIZE
, 8)
96 defframe
(PARAM_DST
, 4)
101 PROLOGUE
(mpn_divrem_1c
)
104 movl PARAM_SIZE, %ecx
105 pushl %edi FRAME_pushl()
108 pushl %esi FRAME_pushl()
110 movl PARAM_DIVISOR, %esi
111 pushl %ebx FRAME_pushl()
114 pushl %ebp FRAME_pushl()
116 movl PARAM_XSIZE, %ebp
119 movl PARAM_CARRY, %edx
122 leal -4(%ebx,%ebp,4), %ebx C dst one limb below integer part
128 PROLOGUE(mpn_divrem_1)
131 movl PARAM_SIZE
, %ecx
132 pushl
%edi FRAME_pushl
()
135 pushl
%esi FRAME_pushl
()
137 movl PARAM_DIVISOR
, %esi
141 pushl
%ebx FRAME_pushl
()
143 movl
-4(%edi,%ecx,4), %eax C src
high limb
147 pushl
%ebp FRAME_pushl
()
149 movl PARAM_XSIZE
, %ebp
152 leal
-4(%ebx,%ebp,4), %ebx C dst one limb below integer part
156 C
high<divisor
, so
high of dst is zero
, and avoid one
div
158 movl
%edx, (%ebx,%ecx,4)
166 C
eax scratch
(quotient
)
169 C
edx scratch
(remainder
)
174 movl
-4(%edi,%ecx,4), %eax
179 movl
%eax, (%ebx,%ecx,4)
192 C
eax scratch
(quotient
)
195 C
edx scratch
(remainder
)
204 movl
%eax, -4(%ebx,%ecx,4)
220 movl PARAM_XSIZE, %ecx
225 cld C better safe than sorry, see mpn/x86/README