1dnl  x86-32 mpn_mod_1_1p, requiring cmov.
2
3dnl  Contributed to the GNU project by Niels Möller and Torbjorn Granlund.
4
5dnl  Copyright 2010, 2011 Free Software Foundation, Inc.
6
7dnl  This file is part of the GNU MP Library.
8dnl
9dnl  The GNU MP Library is free software; you can redistribute it and/or modify
10dnl  it under the terms of either:
11dnl
12dnl    * the GNU Lesser General Public License as published by the Free
13dnl      Software Foundation; either version 3 of the License, or (at your
14dnl      option) any later version.
15dnl
16dnl  or
17dnl
18dnl    * the GNU General Public License as published by the Free Software
19dnl      Foundation; either version 2 of the License, or (at your option) any
20dnl      later version.
21dnl
22dnl  or both in parallel, as here.
23dnl
24dnl  The GNU MP Library is distributed in the hope that it will be useful, but
25dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
26dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
27dnl  for more details.
28dnl
29dnl  You should have received copies of the GNU General Public License and the
30dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
31dnl  see https://www.gnu.org/licenses/.
32
33include(`../config.m4')
34
35C                                 cycles/limb
36C P5                                     ?
37C P6 model 0-8,10-12                     ?
38C P6 model 9  (Banias)                   ?
39C P6 model 13 (Dothan)                   ?
40C P4 model 0  (Willamette)     ?
41C P4 model 1  (?)              ?
42C P4 model 2  (Northwood)      ?
43C P4 model 3  (Prescott)       ?
44C P4 model 4  (Nocona)                   ?
45C AMD K6                       ?
46C AMD K7                       7
47C AMD K8                       ?
48
49define(`B2mb', `%ebx')
50define(`r0', `%esi')
51define(`r2', `%ebp')
52define(`t0', `%edi')
53define(`ap', `%ecx')  C Also shift count
54
55C Stack frame
56C         pre       36(%esp)
57C         b         32(%esp)
58C         n         28(%esp)
59C         ap        24(%esp)
60C         return    20(%esp)
61C         %ebp      16(%esp)
62C         %edi      12(%esp)
63C         %esi      8(%esp)
64C         %ebx      4(%esp)
65C         B2mod     (%esp)
66
67define(`B2modb', `(%esp)')
68define(`n', `28(%esp)')
69define(`b', `32(%esp)')
70define(`pre', `36(%esp)')
71
72C mp_limb_t
73C mpn_mod_1_1p (mp_srcptr ap, mp_size_t n, mp_limb_t b, mp_limb_t pre[4])
74C
75C The pre array contains bi, cnt, B1modb, B2modb
76C Note: This implementation needs B1modb only when cnt > 0
77
78ASM_START()
79          TEXT
80          ALIGN(8)
81PROLOGUE(mpn_mod_1_1p)
82          push      %ebp
83          push      %edi
84          push      %esi
85          push      %ebx
86          mov       32(%esp), %ebp                C pre[]
87
88          mov       12(%ebp), %eax                C B2modb
89          push      %eax                          C Put it on stack
90
91          mov       n, %edx
92          mov       24(%esp), ap
93
94          lea       (ap, %edx, 4), ap
95          mov       -4(ap), %eax
96          cmp       $3, %edx
97          jnc       L(first)
98          mov       -8(ap), r0
99          jmp       L(reduce_two)
100
101L(first):
102          C First iteration, no r2
103          mull      B2modb
104          mov       -12(ap), r0
105          add       %eax, r0
106          mov       -8(ap), %eax
107          adc       %edx, %eax
108          sbb       r2, r2
109          subl      $3, n
110          lea       -16(ap), ap
111          jz        L(reduce_three)
112
113          mov       B2modb, B2mb
114          sub       b, B2mb
115          lea       (B2mb, r0), t0
116          jmp       L(mid)
117
118          ALIGN(16)
119L(top): C Loopmixed to 7 c/l on k7
120          add       %eax, r0
121          lea       (B2mb, r0), t0
122          mov       r2, %eax
123          adc       %edx, %eax
124          sbb       r2, r2
125L(mid):   mull      B2modb
126          and       B2modb, r2
127          add       r0, r2
128          decl      n
129          mov       (ap), r0
130          cmovc(    t0, r2)
131          lea       -4(ap), ap
132          jnz       L(top)
133
134          add       %eax, r0
135          mov       r2, %eax
136          adc       %edx, %eax
137          sbb       r2, r2
138
139L(reduce_three):
140          C Eliminate r2
141          and       b, r2
142          sub       r2, %eax
143
144L(reduce_two):
145          mov       pre, %ebp
146          movb      4(%ebp), %cl
147          test      %cl, %cl
148          jz        L(normalized)
149
150          C Unnormalized, use B1modb to reduce to size < B b
151          mull      8(%ebp)
152          xor       t0, t0
153          add       %eax, r0
154          adc       %edx, t0
155          mov       t0, %eax
156
157          C Left-shift to normalize
158          shld      %cl, r0, %eax C Always use shld?
159
160          shl       %cl, r0
161          jmp       L(udiv)
162
163L(normalized):
164          mov       %eax, t0
165          sub       b, t0
166          cmovnc(   t0, %eax)
167
168L(udiv):
169          lea       1(%eax), t0
170          mull      (%ebp)
171          mov       b, %ebx             C Needed in register for lea
172          add       r0, %eax
173          adc       t0, %edx
174          imul      %ebx, %edx
175          sub       %edx, r0
176          cmp       r0, %eax
177          lea       (%ebx, r0), %eax
178          cmovnc(   r0, %eax)
179          cmp       %ebx, %eax
180          jnc       L(fix)
181L(ok):    shr       %cl, %eax
182
183          add       $4, %esp
184          pop       %ebx
185          pop       %esi
186          pop       %edi
187          pop       %ebp
188
189          ret
190L(fix):   sub       %ebx, %eax
191          jmp       L(ok)
192EPILOGUE()
193
194PROLOGUE(mpn_mod_1_1p_cps)
195          push      %ebp
196          mov       12(%esp), %ebp
197          push      %esi
198          bsr       %ebp, %ecx
199          push      %ebx
200          xor       $31, %ecx
201          mov       16(%esp), %esi
202          sal       %cl, %ebp
203          mov       %ebp, %edx
204          not       %edx
205          mov       $-1, %eax
206          div       %ebp                          C On K7, invert_limb would be a few cycles faster.
207          mov       %eax, (%esi)                  C store bi
208          mov       %ecx, 4(%esi)                 C store cnt
209          neg       %ebp
210          mov       $1, %edx
211          shld      %cl, %eax, %edx
212          imul      %ebp, %edx
213          shr       %cl, %edx
214          imul      %ebp, %eax
215          mov       %edx, 8(%esi)                 C store B1modb
216          mov       %eax, 12(%esi)                C store B2modb
217          pop       %ebx
218          pop       %esi
219          pop       %ebp
220          ret
221EPILOGUE()
222