|
Packit |
5c3484 |
dnl AMD64 mpn_mul_2 -- Multiply an n-limb vector with a 2-limb vector and
|
|
Packit |
5c3484 |
dnl store the result in a third limb vector.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2008, 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 cycles/limb
|
|
Packit |
5c3484 |
C AMD K8,K9 2.275
|
|
Packit |
5c3484 |
C AMD K10 2.275
|
|
Packit |
5c3484 |
C Intel P4 13.5
|
|
Packit |
5c3484 |
C Intel core2 4.0
|
|
Packit |
5c3484 |
C Intel corei 3.8
|
|
Packit |
5c3484 |
C Intel atom ?
|
|
Packit |
5c3484 |
C VIA nano ?
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C This code is the result of running a code generation and optimization tool
|
|
Packit |
5c3484 |
C suite written by David Harvey and Torbjorn Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C TODO
|
|
Packit |
5c3484 |
C * Work on feed-in and wind-down code.
|
|
Packit |
5c3484 |
C * Convert "mov $0" to "xor".
|
|
Packit |
5c3484 |
C * Adjust initial lea to save some bytes.
|
|
Packit |
5c3484 |
C * Perhaps adjust n from n_param&3 value?
|
|
Packit |
5c3484 |
C * Replace with 2.25 c/l sequence.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C INPUT PARAMETERS
|
|
Packit |
5c3484 |
define(`rp', `%rdi')
|
|
Packit |
5c3484 |
define(`up', `%rsi')
|
|
Packit |
5c3484 |
define(`n_param',`%rdx')
|
|
Packit |
5c3484 |
define(`vp', `%rcx')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`v0', `%r8')
|
|
Packit |
5c3484 |
define(`v1', `%r9')
|
|
Packit |
5c3484 |
define(`w0', `%rbx')
|
|
Packit |
5c3484 |
define(`w1', `%rcx')
|
|
Packit |
5c3484 |
define(`w2', `%rbp')
|
|
Packit |
5c3484 |
define(`w3', `%r10')
|
|
Packit |
5c3484 |
define(`n', `%r11')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ABI_SUPPORT(DOS64)
|
|
Packit |
5c3484 |
ABI_SUPPORT(STD64)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
PROLOGUE(mpn_mul_2)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
push %rbx
|
|
Packit |
5c3484 |
push %rbp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (vp), v0
|
|
Packit |
5c3484 |
mov 8(vp), v1
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up), %rax
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov n_param, n
|
|
Packit |
5c3484 |
neg n
|
|
Packit |
5c3484 |
lea -8(up,n_param,8), up
|
|
Packit |
5c3484 |
lea -8(rp,n_param,8), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
and $3, R32(n_param)
|
|
Packit |
5c3484 |
jz L(m2p0)
|
|
Packit |
5c3484 |
cmp $2, R32(n_param)
|
|
Packit |
5c3484 |
jc L(m2p1)
|
|
Packit |
5c3484 |
jz L(m2p2)
|
|
Packit |
5c3484 |
L(m2p3):
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
xor R32(w3), R32(w3)
|
|
Packit |
5c3484 |
mov %rax, w1
|
|
Packit |
5c3484 |
mov %rdx, w2
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
add $-1, n
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add %rax, w2
|
|
Packit |
5c3484 |
jmp L(m23)
|
|
Packit |
5c3484 |
L(m2p0):
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
xor R32(w2), R32(w2)
|
|
Packit |
5c3484 |
mov %rax, w0
|
|
Packit |
5c3484 |
mov %rdx, w1
|
|
Packit |
5c3484 |
jmp L(m20)
|
|
Packit |
5c3484 |
L(m2p1):
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
xor R32(w3), R32(w3)
|
|
Packit |
5c3484 |
xor R32(w0), R32(w0)
|
|
Packit |
5c3484 |
xor R32(w1), R32(w1)
|
|
Packit |
5c3484 |
add $1, n
|
|
Packit |
5c3484 |
jmp L(m2top)
|
|
Packit |
5c3484 |
L(m2p2):
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
xor R32(w0), R32(w0)
|
|
Packit |
5c3484 |
xor R32(w1), R32(w1)
|
|
Packit |
5c3484 |
mov %rax, w2
|
|
Packit |
5c3484 |
mov %rdx, w3
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
add $-2, n
|
|
Packit |
5c3484 |
jmp L(m22)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(32)
|
|
Packit |
5c3484 |
L(m2top):
|
|
Packit |
5c3484 |
add %rax, w3
|
|
Packit |
5c3484 |
adc %rdx, w0
|
|
Packit |
5c3484 |
mov 0(up,n,8), %rax
|
|
Packit |
5c3484 |
adc $0, R32(w1)
|
|
Packit |
5c3484 |
mov $0, R32(w2)
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add %rax, w0
|
|
Packit |
5c3484 |
mov w3, 0(rp,n,8)
|
|
Packit |
5c3484 |
adc %rdx, w1
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
add %rax, w0
|
|
Packit |
5c3484 |
adc %rdx, w1
|
|
Packit |
5c3484 |
adc $0, R32(w2)
|
|
Packit |
5c3484 |
L(m20): mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add %rax, w1
|
|
Packit |
5c3484 |
adc %rdx, w2
|
|
Packit |
5c3484 |
mov 16(up,n,8), %rax
|
|
Packit |
5c3484 |
mov $0, R32(w3)
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
add %rax, w1
|
|
Packit |
5c3484 |
mov 16(up,n,8), %rax
|
|
Packit |
5c3484 |
adc %rdx, w2
|
|
Packit |
5c3484 |
adc $0, R32(w3)
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add %rax, w2
|
|
Packit |
5c3484 |
mov w0, 8(rp,n,8)
|
|
Packit |
5c3484 |
L(m23): adc %rdx, w3
|
|
Packit |
5c3484 |
mov 24(up,n,8), %rax
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
mov $0, R32(w0)
|
|
Packit |
5c3484 |
add %rax, w2
|
|
Packit |
5c3484 |
adc %rdx, w3
|
|
Packit |
5c3484 |
mov w1, 16(rp,n,8)
|
|
Packit |
5c3484 |
mov 24(up,n,8), %rax
|
|
Packit |
5c3484 |
mov $0, R32(w1)
|
|
Packit |
5c3484 |
adc $0, R32(w0)
|
|
Packit |
5c3484 |
L(m22): mul v1
|
|
Packit |
5c3484 |
add %rax, w3
|
|
Packit |
5c3484 |
mov w2, 24(rp,n,8)
|
|
Packit |
5c3484 |
adc %rdx, w0
|
|
Packit |
5c3484 |
mov 32(up,n,8), %rax
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
add $4, n
|
|
Packit |
5c3484 |
js L(m2top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
add %rax, w3
|
|
Packit |
5c3484 |
adc %rdx, w0
|
|
Packit |
5c3484 |
adc $0, R32(w1)
|
|
Packit |
5c3484 |
mov (up), %rax
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
mov w3, (rp)
|
|
Packit |
5c3484 |
add %rax, w0
|
|
Packit |
5c3484 |
adc %rdx, w1
|
|
Packit |
5c3484 |
mov w0, 8(rp)
|
|
Packit |
5c3484 |
mov w1, %rax
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
pop %rbp
|
|
Packit |
5c3484 |
pop %rbx
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|