Home | History | Annotate | Line # | Download | only in powerpc32
aors_n.asm revision 1.1.1.1.2.1
      1 dnl  PowerPC-32 mpn_add_n and mpn_sub_n.
      2 
      3 dnl  Copyright 2002, 2005, 2007 Free Software Foundation, Inc.
      4 
      5 dnl  This file is part of the GNU MP Library.
      6 
      7 dnl  The GNU MP Library is free software; you can redistribute it and/or modify
      8 dnl  it under the terms of the GNU Lesser General Public License as published
      9 dnl  by the Free Software Foundation; either version 3 of the License, or (at
     10 dnl  your option) any later version.
     11 
     12 dnl  The GNU MP Library is distributed in the hope that it will be useful, but
     13 dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
     14 dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
     15 dnl  License for more details.
     16 
     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/.
     19 
     20 include(`../config.m4')
     21 
     22 C                   cycles/limb
     23 C 603e:                  ?
     24 C 604e:                  ?		old: 3.25
     25 C 75x (G3):              ?		old: 3.5
     26 C 7400,7410 (G4):        3.25
     27 C 744x,745x (G4+):       4
     28 C POWER3/PPC630          2
     29 C POWER4/PPC970          2.4
     30 C POWER5                 2.75
     31 C POWER6               40-140
     32 C POWER7                 3
     33 
     34 C INPUT PARAMETERS
     35 define(`rp',	`r3')
     36 define(`up',	`r4')
     37 define(`vp',	`r5')
     38 define(`n',	`r6')
     39 define(`cy',	`r7')
     40 
     41 ifdef(`OPERATION_add_n', `
     42 	define(ADCSBC,	adde)
     43 	define(func,	mpn_add_n)
     44 	define(func_nc,	mpn_add_nc)
     45 	define(IFADD,	`$1')
     46 	define(IFSUB,	`')')
     47 ifdef(`OPERATION_sub_n', `
     48 	define(ADCSBC,	subfe)
     49 	define(func,	mpn_sub_n)
     50 	define(func_nc,	mpn_sub_nc)
     51 	define(IFADD,	`')
     52 	define(IFSUB,	`$1')')
     53 
     54 MULFUNC_PROLOGUE(mpn_add_n mpn_add_nc mpn_sub_n mpn_sub_nc)
     55 
     56 ASM_START()
     57 
     58 PROLOGUE(func_nc)
     59 IFADD(`	addic	r0, cy, -1')		C set carry from argument
     60 IFSUB(`	subfic	r0, cy, 0')		C set carry from argument
     61 	b	L(ent)
     62 EPILOGUE()
     63 
     64 PROLOGUE(func)
     65 IFADD(`	addic	r0, n, 0')		C clear carry
     66 IFSUB(`	addic	r0, n, -1')		C set carry
     67 L(ent):	andi.	r0, n, 3
     68 	addi	r3, r3, -12
     69 	addi	n, n, 1
     70 	cmpwi	cr7, r0, 2
     71 	srwi	r0, n, 2
     72 	sub	r4, r4, r3
     73 	sub	r5, r5, r3
     74 	mtctr	r0
     75 	bne	cr0, L(n00)
     76 
     77 	lwzx	r7, r4, r3		C n = 4, 8, 12, ...
     78 	lwzx	r8, r5, r3
     79 	addi	r3, r3, 4
     80 	lwzx	r9, r4, r3
     81 	ADCSBC	r7, r8, r7
     82 	lwzx	r10, r5, r3
     83 	addi	r3, r3, 4
     84 	b	L(00)
     85 
     86 L(n00):	bge	cr7, L(n01)
     87 	cmpwi	cr0, r0, 0		C n = 1, 5, 9, 13, ...
     88 	lwzx	r0, r4, r3
     89 	lwzx	r6, r5, r3
     90 	addi	r3, r3, 4
     91 	ADCSBC	r0, r6, r0
     92 	ble	L(ret)
     93 L(gt1):	lwzx	r7, r4, r3
     94 	lwzx	r8, r5, r3
     95 	addi	r3, r3, 4
     96 	b	L(01)
     97 
     98 L(n10):
     99 	lwzx	r9, r4, r3		C n = 3, 7, 11, 15, ...
    100 	lwzx	r10, r5, r3
    101 	addi	r3, r3, 4
    102 	lwzx	r11, r4, r3
    103 	ADCSBC	r9, r10, r9
    104 	lwzx	r12, r5, r3
    105 	addi	r3, r3, 4
    106 	b	L(11)
    107 
    108 L(n01):	bne	cr7, L(n10)
    109 	cmpwi	cr0, r0, 0		C n = 2, 6, 10, 14, ...
    110 	lwzx	r11, r4, r3
    111 	lwzx	r12, r5, r3
    112 	addi	r3, r3, 4
    113 	lwzx	r0, r4, r3
    114 	ADCSBC	r11, r12, r11
    115 	lwzx	r6, r5, r3
    116 	addi	r3, r3, 4
    117 	ble	cr0, L(end)
    118 
    119 
    120 L(lp):	lwzx	r7, r4, r3
    121 	ADCSBC	r0, r6, r0
    122 	lwzx	r8, r5, r3
    123 	stwu	r11, 4(r3)
    124 L(01):	lwzx	r9, r4, r3
    125 	ADCSBC	r7, r8, r7
    126 	lwzx	r10, r5, r3
    127 	stwu	r0, 4(r3)
    128 L(00):	lwzx	r11, r4, r3
    129 	ADCSBC	r9, r10, r9
    130 	lwzx	r12, r5, r3
    131 	stwu	r7, 4(r3)
    132 L(11):	lwzx	r0, r4, r3
    133 	ADCSBC	r11, r12, r11
    134 	lwzx	r6, r5, r3
    135 	stwu	r9, 4(r3)
    136 	bdnz	L(lp)
    137 
    138 L(end):	ADCSBC	r0, r6, r0
    139 	stw	r11, 4(r3)
    140 L(ret):	stw	r0, 8(r3)
    141 IFADD(`	li	r3, 0	')
    142 IFADD(`	addze	r3, r3	')
    143 IFSUB(`	subfe	r3, r0, r0')
    144 IFSUB(`	neg	r3, r3')
    145 	blr
    146 EPILOGUE()
    147