Blame mpn/s390_64/copyd.asm

Packit 5c3484
dnl  S/390-64 mpn_copyd
Packit 5c3484
Packit 5c3484
dnl  Copyright 2011 Free Software Foundation, Inc.
Packit 5c3484
Packit 5c3484
dnl  This file is part of the GNU MP Library.
Packit 5c3484
dnl
Packit 5c3484
dnl  The GNU MP Library is free software; you can redistribute it and/or modify
Packit 5c3484
dnl  it under the terms of either:
Packit 5c3484
dnl
Packit 5c3484
dnl    * the GNU Lesser General Public License as published by the Free
Packit 5c3484
dnl      Software Foundation; either version 3 of the License, or (at your
Packit 5c3484
dnl      option) any later version.
Packit 5c3484
dnl
Packit 5c3484
dnl  or
Packit 5c3484
dnl
Packit 5c3484
dnl    * the GNU General Public License as published by the Free Software
Packit 5c3484
dnl      Foundation; either version 2 of the License, or (at your option) any
Packit 5c3484
dnl      later version.
Packit 5c3484
dnl
Packit 5c3484
dnl  or both in parallel, as here.
Packit 5c3484
dnl
Packit 5c3484
dnl  The GNU MP Library is distributed in the hope that it will be useful, but
Packit 5c3484
dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
Packit 5c3484
dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
Packit 5c3484
dnl  for more details.
Packit 5c3484
dnl
Packit 5c3484
dnl  You should have received copies of the GNU General Public License and the
Packit 5c3484
dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
Packit 5c3484
dnl  see https://www.gnu.org/licenses/.
Packit 5c3484
Packit 5c3484
Packit 5c3484
include(`../config.m4')
Packit 5c3484
Packit 5c3484
C            cycles/limb
Packit 5c3484
C z900		 2.67
Packit 5c3484
C z990           1.5
Packit 5c3484
C z9		 ?
Packit 5c3484
C z10		 1.8
Packit 5c3484
C z196		 ?
Packit 5c3484
Packit 5c3484
C FIXME:
Packit 5c3484
C  * Avoid saving/restoring callee-saves registers for n < 3.  This could be
Packit 5c3484
C    done by setting rp=r1, up=r2, i=r0 and r3,r4,r5 for clock regs.
Packit 5c3484
C    We could then use r3...r10 in main loop.
Packit 5c3484
C  * Could we use some EX trick, modifying lmg/stmg, for the feed-in code?
Packit 5c3484
Packit 5c3484
C INPUT PARAMETERS
Packit 5c3484
define(`rp_param',	`%r2')
Packit 5c3484
define(`up_param',	`%r3')
Packit 5c3484
define(`n',		`%r4')
Packit 5c3484
Packit 5c3484
define(`rp',	`%r8')
Packit 5c3484
define(`up',	`%r9')
Packit 5c3484
Packit 5c3484
ASM_START()
Packit 5c3484
PROLOGUE(mpn_copyd)
Packit 5c3484
	stmg	%r6, %r11, 48(%r15)
Packit 5c3484
Packit 5c3484
	sllg	%r1, n, 3
Packit 5c3484
	la	%r10, 8(n)
Packit 5c3484
	aghi	%r1, -64
Packit 5c3484
	srlg	%r10, %r10, 3
Packit 5c3484
	lghi	%r11, -64
Packit 5c3484
Packit 5c3484
	la	rp, 0(%r1,rp_param)	C FIXME use lay on z990 and later
Packit 5c3484
	la	up, 0(%r1,up_param)	C FIXME use lay on z990 and later
Packit 5c3484
Packit 5c3484
	lghi	%r7, 7
Packit 5c3484
	ngr	%r7, n			C n mod 8
Packit 5c3484
	cghi	%r7, 2
Packit 5c3484
	jh	L(b34567)
Packit 5c3484
	cghi	%r7, 1
Packit 5c3484
	je	L(b1)
Packit 5c3484
	jh	L(b2)
Packit 5c3484
Packit 5c3484
L(b0):	brctg	%r10, L(top)
Packit 5c3484
	j	L(end)
Packit 5c3484
Packit 5c3484
L(b1):	lg	%r0, 56(up)
Packit 5c3484
	aghi	up, -8
Packit 5c3484
	stg	%r0, 56(rp)
Packit 5c3484
	aghi	rp, -8
Packit 5c3484
	brctg	%r10, L(top)
Packit 5c3484
	j	L(end)
Packit 5c3484
Packit 5c3484
L(b2):	lmg	%r0, %r1, 48(up)
Packit 5c3484
	aghi	up, -16
Packit 5c3484
	stmg	%r0, %r1, 48(rp)
Packit 5c3484
	aghi	rp, -16
Packit 5c3484
	brctg	%r10, L(top)
Packit 5c3484
	j	L(end)
Packit 5c3484
Packit 5c3484
L(b34567):
Packit 5c3484
	cghi	%r7, 4
Packit 5c3484
	jl	L(b3)
Packit 5c3484
	je	L(b4)
Packit 5c3484
	cghi	%r7, 6
Packit 5c3484
	je	L(b6)
Packit 5c3484
	jh	L(b7)
Packit 5c3484
Packit 5c3484
L(b5):	lmg	%r0, %r4, 24(up)
Packit 5c3484
	aghi	up, -40
Packit 5c3484
	stmg	%r0, %r4, 24(rp)
Packit 5c3484
	aghi	rp, -40
Packit 5c3484
	brctg	%r10, L(top)
Packit 5c3484
	j	L(end)
Packit 5c3484
Packit 5c3484
L(b3):	lmg	%r0, %r2, 40(up)
Packit 5c3484
	aghi	up, -24
Packit 5c3484
	stmg	%r0, %r2, 40(rp)
Packit 5c3484
	aghi	rp, -24
Packit 5c3484
	brctg	%r10, L(top)
Packit 5c3484
	j	L(end)
Packit 5c3484
Packit 5c3484
L(b4):	lmg	%r0, %r3, 32(up)
Packit 5c3484
	aghi	up, -32
Packit 5c3484
	stmg	%r0, %r3, 32(rp)
Packit 5c3484
	aghi	rp, -32
Packit 5c3484
	brctg	%r10, L(top)
Packit 5c3484
	j	L(end)
Packit 5c3484
Packit 5c3484
L(b6):	lmg	%r0, %r5, 16(up)
Packit 5c3484
	aghi	up, -48
Packit 5c3484
	stmg	%r0, %r5, 16(rp)
Packit 5c3484
	aghi	rp, -48
Packit 5c3484
	brctg	%r10, L(top)
Packit 5c3484
	j	L(end)
Packit 5c3484
Packit 5c3484
L(b7):	lmg	%r0, %r6, 8(up)
Packit 5c3484
	aghi	up, -56
Packit 5c3484
	stmg	%r0, %r6, 8(rp)
Packit 5c3484
	aghi	rp, -56
Packit 5c3484
	brctg	%r10, L(top)
Packit 5c3484
	j	L(end)
Packit 5c3484
Packit 5c3484
L(top):	lmg	%r0, %r7, 0(up)
Packit 5c3484
	la	up, 0(%r11,up)
Packit 5c3484
	stmg	%r0, %r7, 0(rp)
Packit 5c3484
	la	rp, 0(%r11,rp)
Packit 5c3484
	brctg	%r10, L(top)
Packit 5c3484
Packit 5c3484
L(end):	lmg	%r6, %r11, 48(%r15)
Packit 5c3484
	br	%r14
Packit 5c3484
EPILOGUE()