|
Packit |
6c4009 |
/* AMD64 __mpn_mul_1 -- Multiply a limb vector with a limb and store
|
|
Packit |
6c4009 |
the result in a second limb vector.
|
|
Packit |
6c4009 |
Copyright (C) 2003-2018 Free Software Foundation, Inc.
|
|
Packit |
6c4009 |
This file is part of the GNU MP Library.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU MP Library is free software; you can redistribute it and/or modify
|
|
Packit |
6c4009 |
it under the terms of the GNU Lesser General Public License as published by
|
|
Packit |
6c4009 |
the Free Software Foundation; either version 2.1 of the License, or (at your
|
|
Packit |
6c4009 |
option) any later version.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU MP Library is distributed in the hope that it will be useful, but
|
|
Packit |
6c4009 |
WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
|
|
Packit |
6c4009 |
or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public
|
|
Packit |
6c4009 |
License for more details.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
You should have received a copy of the GNU Lesser General Public License
|
|
Packit |
6c4009 |
along with the GNU MP Library; see the file COPYING.LIB. If not,
|
|
Packit |
6c4009 |
see <http://www.gnu.org/licenses/>. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#include <sysdep.h>
|
|
Packit |
6c4009 |
#include "asm-syntax.h"
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#define rp %rdi
|
|
Packit |
6c4009 |
#define up %rsi
|
|
Packit |
6c4009 |
#define n_param %rdx
|
|
Packit |
6c4009 |
#define vl %rcx
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#define n %r11
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.text
|
|
Packit |
6c4009 |
ENTRY (__mpn_mul_1)
|
|
Packit |
6c4009 |
push %rbx
|
|
Packit |
6c4009 |
cfi_adjust_cfa_offset (8)
|
|
Packit |
6c4009 |
cfi_rel_offset (%rbx, 0)
|
|
Packit |
6c4009 |
xor %r10, %r10
|
|
Packit |
6c4009 |
mov (up), %rax /* read first u limb early */
|
|
Packit |
6c4009 |
mov n_param, %rbx /* move away n from rdx, mul uses it */
|
|
Packit |
6c4009 |
mul vl
|
|
Packit |
6c4009 |
mov %rbx, %r11
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
add %r10, %rax
|
|
Packit |
6c4009 |
adc $0, %rdx
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
and $3, %ebx
|
|
Packit |
6c4009 |
jz L(b0)
|
|
Packit |
6c4009 |
cmp $2, %ebx
|
|
Packit |
6c4009 |
jz L(b2)
|
|
Packit |
6c4009 |
jg L(b3)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
L(b1): dec n
|
|
Packit |
6c4009 |
jne L(gt1)
|
|
Packit |
6c4009 |
mov %rax, (rp)
|
|
Packit |
6c4009 |
jmp L(ret)
|
|
Packit |
6c4009 |
L(gt1): lea 8(up,n,8), up
|
|
Packit |
6c4009 |
lea -8(rp,n,8), rp
|
|
Packit |
6c4009 |
neg n
|
|
Packit |
6c4009 |
xor %r10, %r10
|
|
Packit |
6c4009 |
xor %ebx, %ebx
|
|
Packit |
6c4009 |
mov %rax, %r9
|
|
Packit |
6c4009 |
mov (up,n,8), %rax
|
|
Packit |
6c4009 |
mov %rdx, %r8
|
|
Packit |
6c4009 |
jmp L(L1)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
L(b0): lea (up,n,8), up
|
|
Packit |
6c4009 |
lea -16(rp,n,8), rp
|
|
Packit |
6c4009 |
neg n
|
|
Packit |
6c4009 |
xor %r10, %r10
|
|
Packit |
6c4009 |
mov %rax, %r8
|
|
Packit |
6c4009 |
mov %rdx, %rbx
|
|
Packit |
6c4009 |
jmp L(L0)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
L(b3): lea -8(up,n,8), up
|
|
Packit |
6c4009 |
lea -24(rp,n,8), rp
|
|
Packit |
6c4009 |
neg n
|
|
Packit |
6c4009 |
mov %rax, %rbx
|
|
Packit |
6c4009 |
mov %rdx, %r10
|
|
Packit |
6c4009 |
jmp L(L3)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
L(b2): lea -16(up,n,8), up
|
|
Packit |
6c4009 |
lea -32(rp,n,8), rp
|
|
Packit |
6c4009 |
neg n
|
|
Packit |
6c4009 |
xor %r8, %r8
|
|
Packit |
6c4009 |
xor %ebx, %ebx
|
|
Packit |
6c4009 |
mov %rax, %r10
|
|
Packit |
6c4009 |
mov 24(up,n,8), %rax
|
|
Packit |
6c4009 |
mov %rdx, %r9
|
|
Packit |
6c4009 |
jmp L(L2)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(top): mov %r10, (rp,n,8)
|
|
Packit |
6c4009 |
add %rax, %r9
|
|
Packit |
6c4009 |
mov (up,n,8), %rax
|
|
Packit |
6c4009 |
adc %rdx, %r8
|
|
Packit |
6c4009 |
mov $0, %r10d
|
|
Packit |
6c4009 |
L(L1): mul vl
|
|
Packit |
6c4009 |
mov %r9, 8(rp,n,8)
|
|
Packit |
6c4009 |
add %rax, %r8
|
|
Packit |
6c4009 |
adc %rdx, %rbx
|
|
Packit |
6c4009 |
L(L0): mov 8(up,n,8), %rax
|
|
Packit |
6c4009 |
mul vl
|
|
Packit |
6c4009 |
mov %r8, 16(rp,n,8)
|
|
Packit |
6c4009 |
add %rax, %rbx
|
|
Packit |
6c4009 |
adc %rdx, %r10
|
|
Packit |
6c4009 |
L(L3): mov 16(up,n,8), %rax
|
|
Packit |
6c4009 |
mul vl
|
|
Packit |
6c4009 |
mov %rbx, 24(rp,n,8)
|
|
Packit |
6c4009 |
mov $0, %r8d # zero
|
|
Packit |
6c4009 |
mov %r8, %rbx # zero
|
|
Packit |
6c4009 |
add %rax, %r10
|
|
Packit |
6c4009 |
mov 24(up,n,8), %rax
|
|
Packit |
6c4009 |
mov %r8, %r9 # zero
|
|
Packit |
6c4009 |
adc %rdx, %r9
|
|
Packit |
6c4009 |
L(L2): mul vl
|
|
Packit |
6c4009 |
add $4, n
|
|
Packit |
6c4009 |
js L(top)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
mov %r10, (rp,n,8)
|
|
Packit |
6c4009 |
add %rax, %r9
|
|
Packit |
6c4009 |
adc %r8, %rdx
|
|
Packit |
6c4009 |
mov %r9, 8(rp,n,8)
|
|
Packit |
6c4009 |
add %r8, %rdx
|
|
Packit |
6c4009 |
L(ret): mov %rdx, %rax
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pop %rbx
|
|
Packit |
6c4009 |
cfi_adjust_cfa_offset (-8)
|
|
Packit |
6c4009 |
cfi_restore (%rbx)
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
END (__mpn_mul_1)
|