|
Packit |
5c3484 |
dnl x86 mpn_gcd_1 optimised for AMD K7.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Contributed to the GNU project by by Kevin Ryde. Rehacked by Torbjorn
|
|
Packit |
5c3484 |
dnl Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2000-2002, 2005, 2009, 2011, 2012, 2014, 2015 Free Software
|
|
Packit |
5c3484 |
dnl 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 |
|
|
Packit |
5c3484 |
C cycles/bit (approx)
|
|
Packit |
5c3484 |
C AMD K7 5.31
|
|
Packit |
5c3484 |
C AMD K8,K9 5.33
|
|
Packit |
5c3484 |
C AMD K10 5.30
|
|
Packit |
5c3484 |
C AMD bd1 ?
|
|
Packit |
5c3484 |
C AMD bobcat 7.02
|
|
Packit |
5c3484 |
C Intel P4-2 10.1
|
|
Packit |
5c3484 |
C Intel P4-3/4 10.0
|
|
Packit |
5c3484 |
C Intel P6/13 5.88
|
|
Packit |
5c3484 |
C Intel core2 6.26
|
|
Packit |
5c3484 |
C Intel NHM 6.83
|
|
Packit |
5c3484 |
C Intel SBR 8.50
|
|
Packit |
5c3484 |
C Intel atom 8.90
|
|
Packit |
5c3484 |
C VIA nano ?
|
|
Packit |
5c3484 |
C Numbers measured with: speed -CD -s16-32 -t16 mpn_gcd_1
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C TODO
|
|
Packit |
5c3484 |
C * Tune overhead, this takes 2-3 cycles more than old code when v0 is tiny.
|
|
Packit |
5c3484 |
C * Stream things better through registers, avoiding some copying.
|
|
Packit |
5c3484 |
C * For ELF, avoid putting GOT base in both ebx and esi. Needs special
|
|
Packit |
5c3484 |
C LEA/LEAL or else discrete code here.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C ctz_table[n] is the number of trailing zeros on n, or MAXSHIFT if n==0.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
deflit(MAXSHIFT, 6)
|
|
Packit |
5c3484 |
deflit(MASK, eval((m4_lshift(1,MAXSHIFT))-1))
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
DEF_OBJECT(ctz_table,64)
|
|
Packit |
5c3484 |
.byte MAXSHIFT
|
|
Packit |
5c3484 |
forloop(i,1,MASK,
|
|
Packit |
5c3484 |
` .byte m4_count_trailing_zeros(i)
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
END_OBJECT(ctz_table)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C Threshold of when to call bmod when U is one limb. Should be about
|
|
Packit |
5c3484 |
C (time_in_cycles(bmod_1,1) + call_overhead) / (cycles/bit).
|
|
Packit |
5c3484 |
define(`DIV_THRES_LOG2', 7)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`up', `%edi')
|
|
Packit |
5c3484 |
define(`n', `%esi')
|
|
Packit |
5c3484 |
define(`v0', `%edx')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
PROLOGUE(mpn_gcd_1)
|
|
Packit |
5c3484 |
push %edi
|
|
Packit |
5c3484 |
push %esi
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov 12(%esp), up
|
|
Packit |
5c3484 |
mov 16(%esp), n
|
|
Packit |
5c3484 |
mov 20(%esp), v0
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up), %eax C U low limb
|
|
Packit |
5c3484 |
or v0, %eax C x | y
|
|
Packit |
5c3484 |
mov $-1, %ecx
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(twos):
|
|
Packit |
5c3484 |
inc %ecx
|
|
Packit |
5c3484 |
shr %eax
|
|
Packit |
5c3484 |
jnc L(twos)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
shr %cl, v0
|
|
Packit |
5c3484 |
mov %ecx, %eax C common twos
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(divide_strip_y):
|
|
Packit |
5c3484 |
shr v0
|
|
Packit |
5c3484 |
jnc L(divide_strip_y)
|
|
Packit |
5c3484 |
adc v0, v0
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
push %eax
|
|
Packit |
5c3484 |
push v0
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
cmp $1, n
|
|
Packit |
5c3484 |
jnz L(reduce_nby1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C Both U and V are single limbs, reduce with bmod if u0 >> v0.
|
|
Packit |
5c3484 |
mov (up), %ecx
|
|
Packit |
5c3484 |
mov %ecx, %eax
|
|
Packit |
5c3484 |
shr $DIV_THRES_LOG2, %ecx
|
|
Packit |
5c3484 |
cmp %ecx, v0
|
|
Packit |
5c3484 |
ja L(reduced)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov v0, %esi
|
|
Packit |
5c3484 |
xor %edx, %edx
|
|
Packit |
5c3484 |
div %esi
|
|
Packit |
5c3484 |
mov %edx, %eax
|
|
Packit |
5c3484 |
jmp L(reduced)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(reduce_nby1):
|
|
Packit |
5c3484 |
ifdef(`PIC_WITH_EBX',`dnl
|
|
Packit |
5c3484 |
push %ebx
|
|
Packit |
5c3484 |
add $-4, %esp
|
|
Packit |
5c3484 |
call L(movl_eip_ebx)
|
|
Packit |
5c3484 |
add $_GLOBAL_OFFSET_TABLE_, %ebx
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
push v0 C param 3
|
|
Packit |
5c3484 |
push n C param 2
|
|
Packit |
5c3484 |
push up C param 1
|
|
Packit |
5c3484 |
cmp $BMOD_1_TO_MOD_1_THRESHOLD, n
|
|
Packit |
5c3484 |
jl L(bmod)
|
|
Packit |
5c3484 |
CALL( mpn_mod_1)
|
|
Packit |
5c3484 |
jmp L(called)
|
|
Packit |
5c3484 |
L(bmod):
|
|
Packit |
5c3484 |
CALL( mpn_modexact_1_odd)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(called):
|
|
Packit |
5c3484 |
ifdef(`PIC_WITH_EBX',`dnl
|
|
Packit |
5c3484 |
add $16, %esp C deallocate params
|
|
Packit |
5c3484 |
pop %ebx
|
|
Packit |
5c3484 |
',`
|
|
Packit |
5c3484 |
add $12, %esp C deallocate params
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
L(reduced):
|
|
Packit |
5c3484 |
pop %edx
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
LEAL( ctz_table, %esi)
|
|
Packit |
5c3484 |
test %eax, %eax
|
|
Packit |
5c3484 |
mov %eax, %ecx
|
|
Packit |
5c3484 |
jnz L(mid)
|
|
Packit |
5c3484 |
jmp L(end)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16) C K8 BC P4 NHM SBR
|
|
Packit |
5c3484 |
L(top): cmovc( %ecx, %eax) C if x-y < 0 0
|
|
Packit |
5c3484 |
cmovc( %edi, %edx) C use x,y-x 0
|
|
Packit |
5c3484 |
L(mid): and $MASK, %ecx C 0
|
|
Packit |
5c3484 |
movzbl (%esi,%ecx), %ecx C 1
|
|
Packit |
5c3484 |
jz L(shift_alot) C 1
|
|
Packit |
5c3484 |
shr %cl, %eax C 3
|
|
Packit |
5c3484 |
mov %eax, %edi C 4
|
|
Packit |
5c3484 |
mov %edx, %ecx C 3
|
|
Packit |
5c3484 |
sub %eax, %ecx C 4
|
|
Packit |
5c3484 |
sub %edx, %eax C 4
|
|
Packit |
5c3484 |
jnz L(top) C 5
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): pop %ecx
|
|
Packit |
5c3484 |
mov %edx, %eax
|
|
Packit |
5c3484 |
shl %cl, %eax
|
|
Packit |
5c3484 |
pop %esi
|
|
Packit |
5c3484 |
pop %edi
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(shift_alot):
|
|
Packit |
5c3484 |
shr $MAXSHIFT, %eax
|
|
Packit |
5c3484 |
mov %eax, %ecx
|
|
Packit |
5c3484 |
jmp L(mid)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifdef(`PIC_WITH_EBX',`dnl
|
|
Packit |
5c3484 |
L(movl_eip_ebx):
|
|
Packit |
5c3484 |
mov (%esp), %ebx
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
EPILOGUE()
|
|
Packit |
5c3484 |
ASM_END()
|