|
Packit |
5c3484 |
dnl x86-64 mpn_rsh1add_n/mpn_rsh1sub_n.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Contributed to the GNU project by Torbjorn Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2011, 2012 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 TODO
|
|
Packit |
5c3484 |
C * Schedule loop less. It is now almost surely overscheduled, resulting in
|
|
Packit |
5c3484 |
C large feed-in and wind-down code.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C cycles/limb
|
|
Packit |
5c3484 |
C AMD K8,K9 ?
|
|
Packit |
5c3484 |
C AMD K10 ?
|
|
Packit |
5c3484 |
C Intel P4 ?
|
|
Packit |
5c3484 |
C Intel core2 ?
|
|
Packit |
5c3484 |
C Intel NMH ?
|
|
Packit |
5c3484 |
C Intel SBR ?
|
|
Packit |
5c3484 |
C Intel atom 5.25
|
|
Packit |
5c3484 |
C VIA nano ?
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C INPUT PARAMETERS
|
|
Packit |
5c3484 |
define(`rp',`%rdi')
|
|
Packit |
5c3484 |
define(`up',`%rsi')
|
|
Packit |
5c3484 |
define(`vp',`%rdx')
|
|
Packit |
5c3484 |
define(`n',`%rcx')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifdef(`OPERATION_rsh1add_n', `
|
|
Packit |
5c3484 |
define(ADDSUB, add)
|
|
Packit |
5c3484 |
define(ADCSBB, adc)
|
|
Packit |
5c3484 |
define(func_n, mpn_rsh1add_n)
|
|
Packit |
5c3484 |
define(func_nc, mpn_rsh1add_nc)')
|
|
Packit |
5c3484 |
ifdef(`OPERATION_rsh1sub_n', `
|
|
Packit |
5c3484 |
define(ADDSUB, sub)
|
|
Packit |
5c3484 |
define(ADCSBB, sbb)
|
|
Packit |
5c3484 |
define(func_n, mpn_rsh1sub_n)
|
|
Packit |
5c3484 |
define(func_nc, mpn_rsh1sub_nc)')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ABI_SUPPORT(DOS64)
|
|
Packit |
5c3484 |
ABI_SUPPORT(STD64)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
MULFUNC_PROLOGUE(mpn_rsh1add_n mpn_rsh1sub_n)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
PROLOGUE(func_n)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
push %rbx
|
|
Packit |
5c3484 |
push %rbp
|
|
Packit |
5c3484 |
push %r12
|
|
Packit |
5c3484 |
push %r13
|
|
Packit |
5c3484 |
push %r14
|
|
Packit |
5c3484 |
push %r15
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up), %r15
|
|
Packit |
5c3484 |
ADDSUB (vp), %r15
|
|
Packit |
5c3484 |
sbb R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
xor R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
shr %r15
|
|
Packit |
5c3484 |
adc R32(%rax), R32(%rax) C return value
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov R32(n), R32(%rbp)
|
|
Packit |
5c3484 |
and $3, R32(%rbp)
|
|
Packit |
5c3484 |
jz L(b0)
|
|
Packit |
5c3484 |
cmp $2, R32(%rbp)
|
|
Packit |
5c3484 |
jae L(b23)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b1): dec n
|
|
Packit |
5c3484 |
jnz L(gt1)
|
|
Packit |
5c3484 |
shl $63, %rbx
|
|
Packit |
5c3484 |
add %rbx, %r15
|
|
Packit |
5c3484 |
mov %r15, (rp)
|
|
Packit |
5c3484 |
jmp L(cj1)
|
|
Packit |
5c3484 |
L(gt1): lea 24(up), up
|
|
Packit |
5c3484 |
lea 24(vp), vp
|
|
Packit |
5c3484 |
mov -16(up), %r9
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov -8(up), %r10
|
|
Packit |
5c3484 |
lea 24(rp), rp
|
|
Packit |
5c3484 |
mov (up), %r11
|
|
Packit |
5c3484 |
ADCSBB -16(vp), %r9
|
|
Packit |
5c3484 |
ADCSBB -8(vp), %r10
|
|
Packit |
5c3484 |
mov %r15, %r12
|
|
Packit |
5c3484 |
ADCSBB (vp), %r11
|
|
Packit |
5c3484 |
mov %r9, %r13
|
|
Packit |
5c3484 |
sbb R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov %r11, %r15
|
|
Packit |
5c3484 |
mov %r10, %r14
|
|
Packit |
5c3484 |
shl $63, %r11
|
|
Packit |
5c3484 |
shl $63, %r10
|
|
Packit |
5c3484 |
shl $63, %r9
|
|
Packit |
5c3484 |
or %r9, %r12
|
|
Packit |
5c3484 |
shr %r13
|
|
Packit |
5c3484 |
mov 8(up), %r8
|
|
Packit |
5c3484 |
shr %r14
|
|
Packit |
5c3484 |
or %r10, %r13
|
|
Packit |
5c3484 |
shr %r15
|
|
Packit |
5c3484 |
or %r11, %r14
|
|
Packit |
5c3484 |
sub $4, n
|
|
Packit |
5c3484 |
jz L(cj5)
|
|
Packit |
5c3484 |
L(gt5): mov 16(up), %r9
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov 24(up), %r10
|
|
Packit |
5c3484 |
ADCSBB 8(vp), %r8
|
|
Packit |
5c3484 |
mov %r15, %rbp
|
|
Packit |
5c3484 |
mov 32(up), %r11
|
|
Packit |
5c3484 |
jmp L(lo1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b23): jnz L(b3)
|
|
Packit |
5c3484 |
mov 8(up), %r8
|
|
Packit |
5c3484 |
sub $2, n
|
|
Packit |
5c3484 |
jnz L(gt2)
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
ADCSBB 8(vp), %r8
|
|
Packit |
5c3484 |
mov %r8, %r12
|
|
Packit |
5c3484 |
jmp L(cj2)
|
|
Packit |
5c3484 |
L(gt2): mov 16(up), %r9
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov 24(up), %r10
|
|
Packit |
5c3484 |
ADCSBB 8(vp), %r8
|
|
Packit |
5c3484 |
mov %r15, %rbp
|
|
Packit |
5c3484 |
mov 32(up), %r11
|
|
Packit |
5c3484 |
ADCSBB 16(vp), %r9
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
ADCSBB 24(vp), %r10
|
|
Packit |
5c3484 |
mov %r9, %r13
|
|
Packit |
5c3484 |
ADCSBB 32(vp), %r11
|
|
Packit |
5c3484 |
mov %r8, %r12
|
|
Packit |
5c3484 |
jmp L(lo2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b3): lea 40(up), up
|
|
Packit |
5c3484 |
lea 8(vp), vp
|
|
Packit |
5c3484 |
mov %r15, %r14
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov -32(up), %r11
|
|
Packit |
5c3484 |
ADCSBB 0(vp), %r11
|
|
Packit |
5c3484 |
lea 8(rp), rp
|
|
Packit |
5c3484 |
sbb R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov %r11, %r15
|
|
Packit |
5c3484 |
shl $63, %r11
|
|
Packit |
5c3484 |
mov -24(up), %r8
|
|
Packit |
5c3484 |
shr %r15
|
|
Packit |
5c3484 |
or %r11, %r14
|
|
Packit |
5c3484 |
sub $3, n
|
|
Packit |
5c3484 |
jnz L(gt3)
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
ADCSBB 8(vp), %r8
|
|
Packit |
5c3484 |
jmp L(cj3)
|
|
Packit |
5c3484 |
L(gt3): mov -16(up), %r9
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov -8(up), %r10
|
|
Packit |
5c3484 |
ADCSBB 8(vp), %r8
|
|
Packit |
5c3484 |
mov %r15, %rbp
|
|
Packit |
5c3484 |
mov (up), %r11
|
|
Packit |
5c3484 |
ADCSBB 16(vp), %r9
|
|
Packit |
5c3484 |
ADCSBB 24(vp), %r10
|
|
Packit |
5c3484 |
mov %r8, %r12
|
|
Packit |
5c3484 |
jmp L(lo3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b0): lea 48(up), up
|
|
Packit |
5c3484 |
lea 16(vp), vp
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov -40(up), %r10
|
|
Packit |
5c3484 |
lea 16(rp), rp
|
|
Packit |
5c3484 |
mov -32(up), %r11
|
|
Packit |
5c3484 |
ADCSBB -8(vp), %r10
|
|
Packit |
5c3484 |
mov %r15, %r13
|
|
Packit |
5c3484 |
ADCSBB (vp), %r11
|
|
Packit |
5c3484 |
sbb R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov %r11, %r15
|
|
Packit |
5c3484 |
mov %r10, %r14
|
|
Packit |
5c3484 |
shl $63, %r11
|
|
Packit |
5c3484 |
shl $63, %r10
|
|
Packit |
5c3484 |
mov -24(up), %r8
|
|
Packit |
5c3484 |
shr %r14
|
|
Packit |
5c3484 |
or %r10, %r13
|
|
Packit |
5c3484 |
shr %r15
|
|
Packit |
5c3484 |
or %r11, %r14
|
|
Packit |
5c3484 |
sub $4, n
|
|
Packit |
5c3484 |
jnz L(gt4)
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
ADCSBB 8(vp), %r8
|
|
Packit |
5c3484 |
jmp L(cj4)
|
|
Packit |
5c3484 |
L(gt4): mov -16(up), %r9
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov -8(up), %r10
|
|
Packit |
5c3484 |
ADCSBB 8(vp), %r8
|
|
Packit |
5c3484 |
mov %r15, %rbp
|
|
Packit |
5c3484 |
mov (up), %r11
|
|
Packit |
5c3484 |
ADCSBB 16(vp), %r9
|
|
Packit |
5c3484 |
jmp L(lo0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(8)
|
|
Packit |
5c3484 |
L(top): mov 16(up), %r9
|
|
Packit |
5c3484 |
shr %r14
|
|
Packit |
5c3484 |
or %r10, %r13
|
|
Packit |
5c3484 |
shr %r15
|
|
Packit |
5c3484 |
or %r11, %r14
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov 24(up), %r10
|
|
Packit |
5c3484 |
mov %rbp, (rp)
|
|
Packit |
5c3484 |
ADCSBB 8(vp), %r8
|
|
Packit |
5c3484 |
mov %r15, %rbp
|
|
Packit |
5c3484 |
lea 32(rp), rp
|
|
Packit |
5c3484 |
mov 32(up), %r11
|
|
Packit |
5c3484 |
L(lo1): ADCSBB 16(vp), %r9
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
mov %r12, -24(rp)
|
|
Packit |
5c3484 |
L(lo0): ADCSBB 24(vp), %r10
|
|
Packit |
5c3484 |
mov %r8, %r12
|
|
Packit |
5c3484 |
mov %r13, -16(rp)
|
|
Packit |
5c3484 |
L(lo3): ADCSBB 32(vp), %r11
|
|
Packit |
5c3484 |
mov %r9, %r13
|
|
Packit |
5c3484 |
mov %r14, -8(rp)
|
|
Packit |
5c3484 |
L(lo2): sbb R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
shl $63, %r8
|
|
Packit |
5c3484 |
mov %r11, %r15
|
|
Packit |
5c3484 |
shr %r12
|
|
Packit |
5c3484 |
mov %r10, %r14
|
|
Packit |
5c3484 |
shl $63, %r9
|
|
Packit |
5c3484 |
lea 32(vp), vp
|
|
Packit |
5c3484 |
shl $63, %r10
|
|
Packit |
5c3484 |
or %r8, %rbp
|
|
Packit |
5c3484 |
shl $63, %r11
|
|
Packit |
5c3484 |
or %r9, %r12
|
|
Packit |
5c3484 |
shr %r13
|
|
Packit |
5c3484 |
mov 8(up), %r8
|
|
Packit |
5c3484 |
sub $4, n
|
|
Packit |
5c3484 |
jg L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): shr %r14
|
|
Packit |
5c3484 |
or %r10, %r13
|
|
Packit |
5c3484 |
shr %r15
|
|
Packit |
5c3484 |
or %r11, %r14
|
|
Packit |
5c3484 |
mov %rbp, (rp)
|
|
Packit |
5c3484 |
lea 32(rp), rp
|
|
Packit |
5c3484 |
L(cj5): add R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
ADCSBB 8(vp), %r8
|
|
Packit |
5c3484 |
mov %r12, -24(rp)
|
|
Packit |
5c3484 |
L(cj4): mov %r13, -16(rp)
|
|
Packit |
5c3484 |
L(cj3): mov %r8, %r12
|
|
Packit |
5c3484 |
mov %r14, -8(rp)
|
|
Packit |
5c3484 |
L(cj2): sbb R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
shl $63, %r8
|
|
Packit |
5c3484 |
shr %r12
|
|
Packit |
5c3484 |
or %r8, %r15
|
|
Packit |
5c3484 |
shl $63, %rbx
|
|
Packit |
5c3484 |
add %rbx, %r12
|
|
Packit |
5c3484 |
mov %r15, (rp)
|
|
Packit |
5c3484 |
mov %r12, 8(rp)
|
|
Packit |
5c3484 |
L(cj1): pop %r15
|
|
Packit |
5c3484 |
pop %r14
|
|
Packit |
5c3484 |
pop %r13
|
|
Packit |
5c3484 |
pop %r12
|
|
Packit |
5c3484 |
pop %rbp
|
|
Packit |
5c3484 |
pop %rbx
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|