Blame sysdeps/alpha/alphaev5/rshift.S

Packit 6c4009
 # Alpha EV5 __mpn_rshift --
Packit 6c4009
Packit 6c4009
 # Copyright (C) 1994-2018 Free Software Foundation, Inc.
Packit 6c4009
Packit 6c4009
 # This file is part of the GNU MP Library.
Packit 6c4009
Packit 6c4009
 # The GNU MP Library is free software; you can redistribute it and/or modify
Packit 6c4009
 # it under the terms of the GNU Lesser General Public License as published by
Packit 6c4009
 # the Free Software Foundation; either version 2.1 of the License, or (at your
Packit 6c4009
 # option) any later version.
Packit 6c4009
Packit 6c4009
 # The GNU MP Library is distributed in the hope that it will be useful, but
Packit 6c4009
 # WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
Packit 6c4009
 # or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
Packit 6c4009
 # License for more details.
Packit 6c4009
Packit 6c4009
 # You should have received a copy of the GNU Lesser General Public License
Packit 6c4009
 # along with the GNU MP Library.  If not, see <http://www.gnu.org/licenses/>.
Packit 6c4009
Packit 6c4009
Packit 6c4009
 # INPUT PARAMETERS
Packit 6c4009
 # res_ptr	r16
Packit 6c4009
 # s1_ptr	r17
Packit 6c4009
 # size		r18
Packit 6c4009
 # cnt		r19
Packit 6c4009
Packit 6c4009
 # This code runs at 3.25 cycles/limb on the EV5.
Packit 6c4009
Packit 6c4009
	.set	noreorder
Packit 6c4009
	.set	noat
Packit 6c4009
.text
Packit 6c4009
	.align	3
Packit 6c4009
	.globl	__mpn_rshift
Packit 6c4009
	.ent	__mpn_rshift
Packit 6c4009
__mpn_rshift:
Packit 6c4009
	.frame	$30,0,$26,0
Packit 6c4009
Packit 6c4009
	ldq	$4,0($17)	# load first limb
Packit 6c4009
	subq	$31,$19,$20
Packit 6c4009
	subq	$18,1,$18
Packit 6c4009
	and	$18,4-1,$28	# number of limbs in first loop
Packit 6c4009
	sll	$4,$20,$0	# compute function result
Packit 6c4009
Packit 6c4009
	beq	$28,.L0
Packit 6c4009
	subq	$18,$28,$18
Packit 6c4009
Packit 6c4009
	.align	3
Packit 6c4009
.Loop0:	ldq	$3,8($17)
Packit 6c4009
	addq	$16,8,$16
Packit 6c4009
	srl	$4,$19,$5
Packit 6c4009
	addq	$17,8,$17
Packit 6c4009
	subq	$28,1,$28
Packit 6c4009
	sll	$3,$20,$6
Packit 6c4009
	or	$3,$3,$4
Packit 6c4009
	or	$5,$6,$8
Packit 6c4009
	stq	$8,-8($16)
Packit 6c4009
	bne	$28,.Loop0
Packit 6c4009
Packit 6c4009
.L0:	srl	$4,$19,$24
Packit 6c4009
	beq	$18,.Lend
Packit 6c4009
 # warm up phase 1
Packit 6c4009
	ldq	$1,8($17)
Packit 6c4009
	subq	$18,4,$18
Packit 6c4009
	ldq	$2,16($17)
Packit 6c4009
	ldq	$3,24($17)
Packit 6c4009
	ldq	$4,32($17)
Packit 6c4009
	beq	$18,.Lend1
Packit 6c4009
 # warm up phase 2
Packit 6c4009
	sll	$1,$20,$7
Packit 6c4009
	srl	$1,$19,$21
Packit 6c4009
	sll	$2,$20,$8
Packit 6c4009
	ldq	$1,40($17)
Packit 6c4009
	srl	$2,$19,$22
Packit 6c4009
	ldq	$2,48($17)
Packit 6c4009
	sll	$3,$20,$5
Packit 6c4009
	or	$7,$24,$7
Packit 6c4009
	srl	$3,$19,$23
Packit 6c4009
	or	$8,$21,$8
Packit 6c4009
	sll	$4,$20,$6
Packit 6c4009
	ldq	$3,56($17)
Packit 6c4009
	srl	$4,$19,$24
Packit 6c4009
	ldq	$4,64($17)
Packit 6c4009
	subq	$18,4,$18
Packit 6c4009
	beq	$18,.Lend2
Packit 6c4009
	.align  4
Packit 6c4009
 # main loop
Packit 6c4009
.Loop:	stq	$7,0($16)
Packit 6c4009
	or	$5,$22,$5
Packit 6c4009
	stq	$8,8($16)
Packit 6c4009
	or	$6,$23,$6
Packit 6c4009
Packit 6c4009
	sll	$1,$20,$7
Packit 6c4009
	subq	$18,4,$18
Packit 6c4009
	srl	$1,$19,$21
Packit 6c4009
	unop	# ldq	$31,-96($17)
Packit 6c4009
Packit 6c4009
	sll	$2,$20,$8
Packit 6c4009
	ldq	$1,72($17)
Packit 6c4009
	srl	$2,$19,$22
Packit 6c4009
	ldq	$2,80($17)
Packit 6c4009
Packit 6c4009
	stq	$5,16($16)
Packit 6c4009
	or	$7,$24,$7
Packit 6c4009
	stq	$6,24($16)
Packit 6c4009
	or	$8,$21,$8
Packit 6c4009
Packit 6c4009
	sll	$3,$20,$5
Packit 6c4009
	unop	# ldq	$31,-96($17)
Packit 6c4009
	srl	$3,$19,$23
Packit 6c4009
	addq	$16,32,$16
Packit 6c4009
Packit 6c4009
	sll	$4,$20,$6
Packit 6c4009
	ldq	$3,88($17)
Packit 6c4009
	srl	$4,$19,$24
Packit 6c4009
	ldq	$4,96($17)
Packit 6c4009
Packit 6c4009
	addq	$17,32,$17
Packit 6c4009
	bne	$18,.Loop
Packit 6c4009
 # cool down phase 2/1
Packit 6c4009
.Lend2:	stq	$7,0($16)
Packit 6c4009
	or	$5,$22,$5
Packit 6c4009
	stq	$8,8($16)
Packit 6c4009
	or	$6,$23,$6
Packit 6c4009
	sll	$1,$20,$7
Packit 6c4009
	srl	$1,$19,$21
Packit 6c4009
	sll	$2,$20,$8
Packit 6c4009
	srl	$2,$19,$22
Packit 6c4009
	stq	$5,16($16)
Packit 6c4009
	or	$7,$24,$7
Packit 6c4009
	stq	$6,24($16)
Packit 6c4009
	or	$8,$21,$8
Packit 6c4009
	sll	$3,$20,$5
Packit 6c4009
	srl	$3,$19,$23
Packit 6c4009
	sll	$4,$20,$6
Packit 6c4009
	srl	$4,$19,$24
Packit 6c4009
 # cool down phase 2/2
Packit 6c4009
	stq	$7,32($16)
Packit 6c4009
	or	$5,$22,$5
Packit 6c4009
	stq	$8,40($16)
Packit 6c4009
	or	$6,$23,$6
Packit 6c4009
	stq	$5,48($16)
Packit 6c4009
	stq	$6,56($16)
Packit 6c4009
 # cool down phase 2/3
Packit 6c4009
	stq	$24,64($16)
Packit 6c4009
	ret	$31,($26),1
Packit 6c4009
Packit 6c4009
 # cool down phase 1/1
Packit 6c4009
.Lend1:	sll	$1,$20,$7
Packit 6c4009
	srl	$1,$19,$21
Packit 6c4009
	sll	$2,$20,$8
Packit 6c4009
	srl	$2,$19,$22
Packit 6c4009
	sll	$3,$20,$5
Packit 6c4009
	or	$7,$24,$7
Packit 6c4009
	srl	$3,$19,$23
Packit 6c4009
	or	$8,$21,$8
Packit 6c4009
	sll	$4,$20,$6
Packit 6c4009
	srl	$4,$19,$24
Packit 6c4009
 # cool down phase 1/2
Packit 6c4009
	stq	$7,0($16)
Packit 6c4009
	or	$5,$22,$5
Packit 6c4009
	stq	$8,8($16)
Packit 6c4009
	or	$6,$23,$6
Packit 6c4009
	stq	$5,16($16)
Packit 6c4009
	stq	$6,24($16)
Packit 6c4009
	stq	$24,32($16)
Packit 6c4009
	ret	$31,($26),1
Packit 6c4009
Packit 6c4009
.Lend:	stq	$24,0($16)
Packit 6c4009
	ret	$31,($26),1
Packit 6c4009
	.end	__mpn_rshift