1dnl  Intel Atom mpn_and_n,...,mpn_xnor_n -- bitwise logical operations.
2
3dnl  Copyright 2011 Free Software Foundation, Inc.
4
5dnl  Contributed to the GNU project by Marco Bodrato.
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				op	nop	opn
37C P5
38C P6 model 0-8,10-12
39C P6 model 9  (Banias)
40C P6 model 13 (Dothan)
41C P4 model 0  (Willamette)
42C P4 model 1  (?)
43C P4 model 2  (Northwood)
44C P4 model 3  (Prescott)
45C P4 model 4  (Nocona)
46C Intel Atom			 3	 3.5	 3.5
47C AMD K6
48C AMD K7
49C AMD K8
50C AMD K10
51
52define(M4_choose_op,
53`ifdef(`OPERATION_$1',`
54define(`M4_function', `mpn_$1')
55define(`M4_want_pre', `$4')
56define(`M4_inst',     `$3')
57define(`M4_want_post',`$2')
58')')
59define(M4pre, `ifelse(M4_want_pre, yes,`$1')')
60define(M4post,`ifelse(M4_want_post,yes,`$1')')
61
62M4_choose_op( and_n,     , andl,    )
63M4_choose_op( andn_n,    , andl, yes)
64M4_choose_op( nand_n, yes, andl,    )
65M4_choose_op( ior_n,     ,  orl,    )
66M4_choose_op( iorn_n,    ,  orl, yes)
67M4_choose_op( nior_n, yes,  orl,    )
68M4_choose_op( xor_n,     , xorl,    )
69M4_choose_op( xnor_n, yes, xorl,    )
70
71ifdef(`M4_function',,
72`m4_error(`Unrecognised or undefined OPERATION symbol
73')')
74
75MULFUNC_PROLOGUE(mpn_and_n mpn_andn_n mpn_nand_n mpn_ior_n mpn_iorn_n mpn_nior_n mpn_xor_n mpn_xnor_n)
76
77C void M4_function (mp_ptr dst, mp_srcptr src2, mp_srcptr src1, mp_size_t size);
78C
79
80defframe(PARAM_SIZE, 16)
81defframe(PARAM_SRC1, 12)
82defframe(PARAM_SRC2, 8)
83defframe(PARAM_DST,  4)
84
85dnl  re-use parameter space
86define(SAVE_RP,`PARAM_SIZE')
87define(SAVE_VP,`PARAM_SRC1')
88define(SAVE_UP,`PARAM_DST')
89
90define(`rp',  `%edi')
91define(`up',  `%esi')
92define(`vp',  `%ebx')
93define(`cnt', `%eax')
94define(`r1',  `%ecx')
95define(`r2',  `%edx')
96
97ASM_START()
98	TEXT
99	ALIGN(16)
100deflit(`FRAME',0)
101
102PROLOGUE(M4_function)
103	mov	PARAM_SIZE, cnt		C size
104	mov	rp, SAVE_RP
105	mov	PARAM_DST, rp
106	mov	up, SAVE_UP
107	mov	PARAM_SRC1, up
108	shr	cnt			C size >> 1
109	mov	vp, SAVE_VP
110	mov	PARAM_SRC2, vp
111	mov	(up), r1
112	jz	L(end)			C size == 1
113	jnc	L(even)			C size % 2 == 0
114
115	ALIGN(16)
116L(oop):
117M4pre(`	notl_or_xorl_GMP_NUMB_MASK(r1)')
118	M4_inst	(vp), r1
119	lea	8(up), up
120	mov	-4(up), r2
121M4post(`	notl_or_xorl_GMP_NUMB_MASK(r1)')
122	lea	8(vp), vp
123	mov	r1, (rp)
124L(entry):
125M4pre(`	notl_or_xorl_GMP_NUMB_MASK(r2)')
126	M4_inst	-4(vp), r2
127	lea	8(rp), rp
128M4post(`	notl_or_xorl_GMP_NUMB_MASK(r2)')
129	dec	cnt
130	mov	(up), r1
131	mov	r2, -4(rp)
132	jnz	L(oop)
133
134L(end):
135M4pre(`	notl_or_xorl_GMP_NUMB_MASK(r1)')
136	mov	SAVE_UP, up
137	M4_inst	(vp), r1
138M4post(`notl_or_xorl_GMP_NUMB_MASK(r1)')
139	mov	SAVE_VP, vp
140	mov	r1, (rp)
141	mov	SAVE_RP, rp
142	ret
143
144L(even):
145	mov	r1, r2
146	lea	4(up), up
147	lea	4(vp), vp
148	lea	-4(rp), rp
149	jmp	L(entry)
150EPILOGUE()
151ASM_END()
152