diff options
author | Samuel Thibault <samuel.thibault@ens-lyon.org> | 2016-08-20 19:50:45 +0200 |
---|---|---|
committer | Samuel Thibault <samuel.thibault@ens-lyon.org> | 2016-08-20 19:50:45 +0200 |
commit | 4dd9e35bfd35d3138bc44169baba098005bad51e (patch) | |
tree | a4939c43a9c3fe00eb27f023e14acc5e1fe8808c /sysdeps/alpha/lshift.S | |
parent | bd42a4599d1b6f77bcfe1e4f67b7cbd9e1cb2dfd (diff) | |
parent | f76453c31593957fec1a99b986bfa5506618b79c (diff) |
Merge commit 'refs/top-bases/t/bigmem' into t/bigmem
Diffstat (limited to 'sysdeps/alpha/lshift.S')
-rw-r--r-- | sysdeps/alpha/lshift.S | 107 |
1 files changed, 107 insertions, 0 deletions
diff --git a/sysdeps/alpha/lshift.S b/sysdeps/alpha/lshift.S new file mode 100644 index 0000000000..8926c8fe88 --- /dev/null +++ b/sysdeps/alpha/lshift.S @@ -0,0 +1,107 @@ + # Alpha 21064 __mpn_lshift -- + + # Copyright (C) 1994-2015 Free Software Foundation, Inc. + + # This file is part of the GNU MP Library. + + # The GNU MP Library is free software; you can redistribute it and/or modify + # it under the terms of the GNU Lesser General Public License as published by + # the Free Software Foundation; either version 2.1 of the License, or (at your + # option) any later version. + + # The GNU MP Library is distributed in the hope that it will be useful, but + # WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY + # or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public + # License for more details. + + # You should have received a copy of the GNU Lesser General Public License + # along with the GNU MP Library. If not, see <http://www.gnu.org/licenses/>. + + + # INPUT PARAMETERS + # res_ptr r16 + # s1_ptr r17 + # size r18 + # cnt r19 + + # This code runs at 4.8 cycles/limb on the 21064. With infinite unrolling, + # it would take 4 cycles/limb. It should be possible to get down to 3 + # cycles/limb since both ldq and stq can be paired with the other used + # instructions. But there are many restrictions in the 21064 pipeline that + # makes it hard, if not impossible, to get down to 3 cycles/limb: + + # 1. ldq has a 3 cycle delay, srl and sll have a 2 cycle delay. + # 2. Only aligned instruction pairs can be paired. + # 3. The store buffer or silo might not be able to deal with the bandwidth. + + .set noreorder + .set noat +.text + .align 3 + .globl __mpn_lshift + .ent __mpn_lshift +__mpn_lshift: + .frame $30,0,$26,0 + + s8addq $18,$17,$17 # make r17 point at end of s1 + ldq $4,-8($17) # load first limb + subq $17,8,$17 + subq $31,$19,$7 + s8addq $18,$16,$16 # make r16 point at end of RES + subq $18,1,$18 + and $18,4-1,$20 # number of limbs in first loop + srl $4,$7,$0 # compute function result + + beq $20,.L0 + subq $18,$20,$18 + + .align 3 +.Loop0: + ldq $3,-8($17) + subq $16,8,$16 + subq $17,8,$17 + subq $20,1,$20 + sll $4,$19,$5 + srl $3,$7,$6 + bis $3,$3,$4 + bis $5,$6,$8 + stq $8,0($16) + bne $20,.Loop0 + +.L0: beq $18,.Lend + + .align 3 +.Loop: ldq $3,-8($17) + subq $16,32,$16 + subq $18,4,$18 + sll $4,$19,$5 + srl $3,$7,$6 + + ldq $4,-16($17) + sll $3,$19,$1 + bis $5,$6,$8 + stq $8,24($16) + srl $4,$7,$2 + + ldq $3,-24($17) + sll $4,$19,$5 + bis $1,$2,$8 + stq $8,16($16) + srl $3,$7,$6 + + ldq $4,-32($17) + sll $3,$19,$1 + bis $5,$6,$8 + stq $8,8($16) + srl $4,$7,$2 + + subq $17,32,$17 + bis $1,$2,$8 + stq $8,0($16) + + bgt $18,.Loop + +.Lend: sll $4,$19,$8 + stq $8,-8($16) + ret $31,($26),1 + .end __mpn_lshift |