|
Packit |
5c3484 |
dnl S/390-64 mpn_lshift.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2011, 2012, 2014 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 |
include(`../config.m4')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C cycles/limb
|
|
Packit |
5c3484 |
C z900 7
|
|
Packit |
5c3484 |
C z990 3
|
|
Packit |
5c3484 |
C z9 ?
|
|
Packit |
5c3484 |
C z10 6
|
|
Packit |
5c3484 |
C z196 ?
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C NOTES
|
|
Packit |
5c3484 |
C * This uses discrete loads and stores in a software pipeline. Using lmg and
|
|
Packit |
5c3484 |
C stmg is not faster.
|
|
Packit |
5c3484 |
C * One could assume more pipelining could approach 2.5 c/l, but we have not
|
|
Packit |
5c3484 |
C found any 8-way loop that runs better than the current 4-way loop.
|
|
Packit |
5c3484 |
C * Consider using the same feed-in code for 1 <= n <= 3 as for n mod 4,
|
|
Packit |
5c3484 |
C similarly to the x86_64 sqr_basecase feed-in.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C INPUT PARAMETERS
|
|
Packit |
5c3484 |
define(`rp', `%r2')
|
|
Packit |
5c3484 |
define(`up', `%r3')
|
|
Packit |
5c3484 |
define(`n', `%r4')
|
|
Packit |
5c3484 |
define(`cnt', `%r5')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`tnc', `%r6')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
PROLOGUE(mpn_lshift)
|
|
Packit |
5c3484 |
cghi n, 3
|
|
Packit |
5c3484 |
jh L(gt1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
stmg %r6, %r7, 48(%r15)
|
|
Packit |
5c3484 |
larl %r1, L(tab)-4
|
|
Packit |
5c3484 |
lcgr tnc, cnt
|
|
Packit |
5c3484 |
sllg n, n, 2
|
|
Packit |
5c3484 |
b 0(n,%r1)
|
|
Packit |
5c3484 |
L(tab): j L(n1)
|
|
Packit |
5c3484 |
j L(n2)
|
|
Packit |
5c3484 |
j L(n3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(n1): lg %r1, 0(up)
|
|
Packit |
5c3484 |
sllg %r0, %r1, 0(cnt)
|
|
Packit |
5c3484 |
stg %r0, 0(rp)
|
|
Packit |
5c3484 |
srlg %r2, %r1, 0(tnc)
|
|
Packit |
5c3484 |
lg %r6, 48(%r15) C restoring r7 not needed
|
|
Packit |
5c3484 |
br %r14
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(n2): lg %r1, 8(up)
|
|
Packit |
5c3484 |
srlg %r4, %r1, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r0, %r1, 0(cnt)
|
|
Packit |
5c3484 |
j L(cj)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(n3): lg %r1, 16(up)
|
|
Packit |
5c3484 |
srlg %r4, %r1, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r0, %r1, 0(cnt)
|
|
Packit |
5c3484 |
lg %r1, 8(up)
|
|
Packit |
5c3484 |
srlg %r7, %r1, 0(tnc)
|
|
Packit |
5c3484 |
ogr %r7, %r0
|
|
Packit |
5c3484 |
sllg %r0, %r1, 0(cnt)
|
|
Packit |
5c3484 |
stg %r7, 16(rp)
|
|
Packit |
5c3484 |
L(cj): lg %r1, 0(up)
|
|
Packit |
5c3484 |
srlg %r7, %r1, 0(tnc)
|
|
Packit |
5c3484 |
ogr %r7, %r0
|
|
Packit |
5c3484 |
sllg %r0, %r1, 0(cnt)
|
|
Packit |
5c3484 |
stg %r7, 8(rp)
|
|
Packit |
5c3484 |
stg %r0, 0(rp)
|
|
Packit |
5c3484 |
lgr %r2, %r4
|
|
Packit |
5c3484 |
lmg %r6, %r7, 48(%r15)
|
|
Packit |
5c3484 |
br %r14
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(gt1): stmg %r6, %r13, 48(%r15)
|
|
Packit |
5c3484 |
lcgr tnc, cnt C tnc = -cnt
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
sllg %r1, n, 3
|
|
Packit |
5c3484 |
srlg %r0, n, 2 C loop count
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
agr up, %r1 C point up at end of U
|
|
Packit |
5c3484 |
agr rp, %r1 C point rp at end of R
|
|
Packit |
5c3484 |
aghi up, -56
|
|
Packit |
5c3484 |
aghi rp, -40
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
lghi %r7, 3
|
|
Packit |
5c3484 |
ngr %r7, n
|
|
Packit |
5c3484 |
je L(b0)
|
|
Packit |
5c3484 |
cghi %r7, 2
|
|
Packit |
5c3484 |
jl L(b1)
|
|
Packit |
5c3484 |
je L(b2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b3): lg %r7, 48(up)
|
|
Packit |
5c3484 |
srlg %r9, %r7, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r11, %r7, 0(cnt)
|
|
Packit |
5c3484 |
lg %r8, 40(up)
|
|
Packit |
5c3484 |
lg %r7, 32(up)
|
|
Packit |
5c3484 |
srlg %r4, %r8, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r13, %r8, 0(cnt)
|
|
Packit |
5c3484 |
ogr %r11, %r4
|
|
Packit |
5c3484 |
la rp, 16(rp)
|
|
Packit |
5c3484 |
j L(lm3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b2): lg %r8, 48(up)
|
|
Packit |
5c3484 |
lg %r7, 40(up)
|
|
Packit |
5c3484 |
srlg %r9, %r8, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r13, %r8, 0(cnt)
|
|
Packit |
5c3484 |
la rp, 24(rp)
|
|
Packit |
5c3484 |
la up, 8(up)
|
|
Packit |
5c3484 |
j L(lm2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b1): lg %r7, 48(up)
|
|
Packit |
5c3484 |
srlg %r9, %r7, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r11, %r7, 0(cnt)
|
|
Packit |
5c3484 |
lg %r8, 40(up)
|
|
Packit |
5c3484 |
lg %r7, 32(up)
|
|
Packit |
5c3484 |
srlg %r4, %r8, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r10, %r8, 0(cnt)
|
|
Packit |
5c3484 |
ogr %r11, %r4
|
|
Packit |
5c3484 |
la rp, 32(rp)
|
|
Packit |
5c3484 |
la up, 16(up)
|
|
Packit |
5c3484 |
j L(lm1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b0): lg %r8, 48(up)
|
|
Packit |
5c3484 |
lg %r7, 40(up)
|
|
Packit |
5c3484 |
srlg %r9, %r8, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r10, %r8, 0(cnt)
|
|
Packit |
5c3484 |
la rp, 40(rp)
|
|
Packit |
5c3484 |
la up, 24(up)
|
|
Packit |
5c3484 |
j L(lm0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(8)
|
|
Packit |
5c3484 |
L(top): srlg %r4, %r8, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r13, %r8, 0(cnt)
|
|
Packit |
5c3484 |
ogr %r11, %r4
|
|
Packit |
5c3484 |
stg %r10, 24(rp)
|
|
Packit |
5c3484 |
L(lm3): stg %r11, 16(rp)
|
|
Packit |
5c3484 |
L(lm2): srlg %r12, %r7, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r11, %r7, 0(cnt)
|
|
Packit |
5c3484 |
lg %r8, 24(up)
|
|
Packit |
5c3484 |
lg %r7, 16(up)
|
|
Packit |
5c3484 |
ogr %r13, %r12
|
|
Packit |
5c3484 |
srlg %r4, %r8, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r10, %r8, 0(cnt)
|
|
Packit |
5c3484 |
ogr %r11, %r4
|
|
Packit |
5c3484 |
stg %r13, 8(rp)
|
|
Packit |
5c3484 |
L(lm1): stg %r11, 0(rp)
|
|
Packit |
5c3484 |
L(lm0): srlg %r12, %r7, 0(tnc)
|
|
Packit |
5c3484 |
aghi rp, -32
|
|
Packit |
5c3484 |
sllg %r11, %r7, 0(cnt)
|
|
Packit |
5c3484 |
lg %r8, 8(up)
|
|
Packit |
5c3484 |
lg %r7, 0(up)
|
|
Packit |
5c3484 |
aghi up, -32
|
|
Packit |
5c3484 |
ogr %r10, %r12
|
|
Packit |
5c3484 |
brctg %r0, L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): srlg %r4, %r8, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r13, %r8, 0(cnt)
|
|
Packit |
5c3484 |
ogr %r11, %r4
|
|
Packit |
5c3484 |
stg %r10, 24(rp)
|
|
Packit |
5c3484 |
stg %r11, 16(rp)
|
|
Packit |
5c3484 |
srlg %r12, %r7, 0(tnc)
|
|
Packit |
5c3484 |
sllg %r11, %r7, 0(cnt)
|
|
Packit |
5c3484 |
ogr %r13, %r12
|
|
Packit |
5c3484 |
stg %r13, 8(rp)
|
|
Packit |
5c3484 |
stg %r11, 0(rp)
|
|
Packit |
5c3484 |
lgr %r2, %r9
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
lmg %r6, %r13, 48(%r15)
|
|
Packit |
5c3484 |
br %r14
|
|
Packit |
5c3484 |
EPILOGUE()
|