|
Packit |
5c3484 |
dnl x86-64 mpn_rsh1add_n/mpn_rsh1sub_n optimized for Pentium 4.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Contributed to the GNU project by Torbjorn Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2007, 2008, 2010-2012 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 |
|
|
Packit |
5c3484 |
C cycles/limb
|
|
Packit |
5c3484 |
C AMD K8,K9 4.13
|
|
Packit |
5c3484 |
C AMD K10 4.13
|
|
Packit |
5c3484 |
C Intel P4 5.70
|
|
Packit |
5c3484 |
C Intel core2 4.75
|
|
Packit |
5c3484 |
C Intel corei 5
|
|
Packit |
5c3484 |
C Intel atom 8.75
|
|
Packit |
5c3484 |
C VIA nano 5.25
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C TODO
|
|
Packit |
5c3484 |
C * Try to make this smaller, 746 bytes seem excessive for this 2nd class
|
|
Packit |
5c3484 |
C function. Less sw pipelining would help, and since we now probably
|
|
Packit |
5c3484 |
C pipeline somewhat too deeply, it might not affect performance too much.
|
|
Packit |
5c3484 |
C * A separate small-n loop might speed things as well as make things smaller.
|
|
Packit |
5c3484 |
C That loop should be selected before pushing registers.
|
|
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 |
define(`cy', `%r8')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifdef(`OPERATION_rsh1add_n', `
|
|
Packit |
5c3484 |
define(ADDSUB, add)
|
|
Packit |
5c3484 |
define(func, 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(func, 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_rsh1add_nc mpn_rsh1sub_n mpn_rsh1sub_nc)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
PROLOGUE(func)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
xor %r8, %r8
|
|
Packit |
5c3484 |
IFDOS(` jmp L(ent) ')
|
|
Packit |
5c3484 |
EPILOGUE()
|
|
Packit |
5c3484 |
PROLOGUE(func_nc)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
IFDOS(` mov 56(%rsp), %r8 ')
|
|
Packit |
5c3484 |
L(ent): push %rbx
|
|
Packit |
5c3484 |
push %r12
|
|
Packit |
5c3484 |
push %r13
|
|
Packit |
5c3484 |
push %r14
|
|
Packit |
5c3484 |
push %r15
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (vp), %r9
|
|
Packit |
5c3484 |
mov (up), %r15
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov R32(n), R32(%rax)
|
|
Packit |
5c3484 |
and $3, R32(%rax)
|
|
Packit |
5c3484 |
jne L(n00)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov R32(%r8), R32(%rbx) C n = 0, 4, 8, ...
|
|
Packit |
5c3484 |
mov 8(up), %r10
|
|
Packit |
5c3484 |
ADDSUB %r9, %r15
|
|
Packit |
5c3484 |
mov 8(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rax)
|
|
Packit |
5c3484 |
ADDSUB %rbx, %r15 C return bit
|
|
Packit |
5c3484 |
jnc 1f
|
|
Packit |
5c3484 |
mov $1, R8(%rax)
|
|
Packit |
5c3484 |
1: mov 16(up), %r12
|
|
Packit |
5c3484 |
ADDSUB %r9, %r10
|
|
Packit |
5c3484 |
mov 16(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
mov %r15, %r13
|
|
Packit |
5c3484 |
ADDSUB %rax, %r10
|
|
Packit |
5c3484 |
jnc 1f
|
|
Packit |
5c3484 |
mov $1, R8(%rbx)
|
|
Packit |
5c3484 |
1: mov 24(up), %r11
|
|
Packit |
5c3484 |
ADDSUB %r9, %r12
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
mov 24(vp), %r9
|
|
Packit |
5c3484 |
lea 32(vp), vp
|
|
Packit |
5c3484 |
setc R8(%rax)
|
|
Packit |
5c3484 |
mov %r10, %r14
|
|
Packit |
5c3484 |
shl $63, %r10
|
|
Packit |
5c3484 |
shr %r13
|
|
Packit |
5c3484 |
jmp L(L00)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(n00): cmp $2, R32(%rax)
|
|
Packit |
5c3484 |
jnc L(n01)
|
|
Packit |
5c3484 |
xor R32(%rbx), R32(%rbx) C n = 1, 5, 9, ...
|
|
Packit |
5c3484 |
lea -24(rp), rp
|
|
Packit |
5c3484 |
mov R32(%r8), R32(%rax)
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
jnz L(gt1)
|
|
Packit |
5c3484 |
ADDSUB %r9, %r15
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
ADDSUB %rax, %r15
|
|
Packit |
5c3484 |
jnc 1f
|
|
Packit |
5c3484 |
mov $1, R8(%rbx)
|
|
Packit |
5c3484 |
1: mov %r15, %r14
|
|
Packit |
5c3484 |
shl $63, %rbx
|
|
Packit |
5c3484 |
shr %r14
|
|
Packit |
5c3484 |
jmp L(cj1)
|
|
Packit |
5c3484 |
L(gt1): mov 8(up), %r8
|
|
Packit |
5c3484 |
ADDSUB %r9, %r15
|
|
Packit |
5c3484 |
mov 8(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
ADDSUB %rax, %r15
|
|
Packit |
5c3484 |
jnc 1f
|
|
Packit |
5c3484 |
mov $1, R8(%rbx)
|
|
Packit |
5c3484 |
1: mov 16(up), %r10
|
|
Packit |
5c3484 |
ADDSUB %r9, %r8
|
|
Packit |
5c3484 |
mov 16(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rax)
|
|
Packit |
5c3484 |
mov %r15, %r14
|
|
Packit |
5c3484 |
ADDSUB %rbx, %r8
|
|
Packit |
5c3484 |
jnc 1f
|
|
Packit |
5c3484 |
mov $1, R8(%rax)
|
|
Packit |
5c3484 |
1: mov 24(up), %r12
|
|
Packit |
5c3484 |
ADDSUB %r9, %r10
|
|
Packit |
5c3484 |
mov 24(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
mov %r8, %r13
|
|
Packit |
5c3484 |
shl $63, %r8
|
|
Packit |
5c3484 |
shr %r14
|
|
Packit |
5c3484 |
lea 8(up), up
|
|
Packit |
5c3484 |
lea 8(vp), vp
|
|
Packit |
5c3484 |
jmp L(L01)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(n01): jne L(n10)
|
|
Packit |
5c3484 |
lea -16(rp), rp C n = 2, 6, 10, ...
|
|
Packit |
5c3484 |
mov R32(%r8), R32(%rbx)
|
|
Packit |
5c3484 |
mov 8(up), %r11
|
|
Packit |
5c3484 |
ADDSUB %r9, %r15
|
|
Packit |
5c3484 |
mov 8(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rax)
|
|
Packit |
5c3484 |
ADDSUB %rbx, %r15
|
|
Packit |
5c3484 |
jnc 1f
|
|
Packit |
5c3484 |
mov $1, R8(%rax)
|
|
Packit |
5c3484 |
1: sub $2, n
|
|
Packit |
5c3484 |
jnz L(gt2)
|
|
Packit |
5c3484 |
ADDSUB %r9, %r11
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
mov %r15, %r13
|
|
Packit |
5c3484 |
ADDSUB %rax, %r11
|
|
Packit |
5c3484 |
jnc 1f
|
|
Packit |
5c3484 |
mov $1, R8(%rbx)
|
|
Packit |
5c3484 |
1: mov %r11, %r14
|
|
Packit |
5c3484 |
shl $63, %r11
|
|
Packit |
5c3484 |
shr %r13
|
|
Packit |
5c3484 |
jmp L(cj2)
|
|
Packit |
5c3484 |
L(gt2): mov 16(up), %r8
|
|
Packit |
5c3484 |
ADDSUB %r9, %r11
|
|
Packit |
5c3484 |
mov 16(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
mov %r15, %r13
|
|
Packit |
5c3484 |
ADDSUB %rax, %r11
|
|
Packit |
5c3484 |
jnc 1f
|
|
Packit |
5c3484 |
mov $1, R8(%rbx)
|
|
Packit |
5c3484 |
1: mov 24(up), %r10
|
|
Packit |
5c3484 |
ADDSUB %r9, %r8
|
|
Packit |
5c3484 |
mov 24(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rax)
|
|
Packit |
5c3484 |
mov %r11, %r14
|
|
Packit |
5c3484 |
shl $63, %r11
|
|
Packit |
5c3484 |
shr %r13
|
|
Packit |
5c3484 |
lea 16(up), up
|
|
Packit |
5c3484 |
lea 16(vp), vp
|
|
Packit |
5c3484 |
jmp L(L10)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(n10): xor R32(%rbx), R32(%rbx) C n = 3, 7, 11, ...
|
|
Packit |
5c3484 |
lea -8(rp), rp
|
|
Packit |
5c3484 |
mov R32(%r8), R32(%rax)
|
|
Packit |
5c3484 |
mov 8(up), %r12
|
|
Packit |
5c3484 |
ADDSUB %r9, %r15
|
|
Packit |
5c3484 |
mov 8(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
ADDSUB %rax, %r15
|
|
Packit |
5c3484 |
jnc 1f
|
|
Packit |
5c3484 |
mov $1, R8(%rbx)
|
|
Packit |
5c3484 |
1: mov 16(up), %r11
|
|
Packit |
5c3484 |
ADDSUB %r9, %r12
|
|
Packit |
5c3484 |
mov 16(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rax)
|
|
Packit |
5c3484 |
mov %r15, %r14
|
|
Packit |
5c3484 |
ADDSUB %rbx, %r12
|
|
Packit |
5c3484 |
jnc 1f
|
|
Packit |
5c3484 |
mov $1, R8(%rax)
|
|
Packit |
5c3484 |
1: sub $3, n
|
|
Packit |
5c3484 |
jnz L(gt3)
|
|
Packit |
5c3484 |
ADDSUB %r9, %r11
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
mov %r12, %r13
|
|
Packit |
5c3484 |
shl $63, %r12
|
|
Packit |
5c3484 |
shr %r14
|
|
Packit |
5c3484 |
jmp L(cj3)
|
|
Packit |
5c3484 |
L(gt3): mov 24(up), %r8
|
|
Packit |
5c3484 |
ADDSUB %r9, %r11
|
|
Packit |
5c3484 |
mov 24(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
mov %r12, %r13
|
|
Packit |
5c3484 |
shl $63, %r12
|
|
Packit |
5c3484 |
shr %r14
|
|
Packit |
5c3484 |
lea 24(up), up
|
|
Packit |
5c3484 |
lea 24(vp), vp
|
|
Packit |
5c3484 |
jmp L(L11)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(c0): mov $1, R8(%rbx)
|
|
Packit |
5c3484 |
jmp L(rc0)
|
|
Packit |
5c3484 |
L(c1): mov $1, R8(%rax)
|
|
Packit |
5c3484 |
jmp L(rc1)
|
|
Packit |
5c3484 |
L(c2): mov $1, R8(%rbx)
|
|
Packit |
5c3484 |
jmp L(rc2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(top): mov (up), %r8 C not on critical path
|
|
Packit |
5c3484 |
or %r13, %r10
|
|
Packit |
5c3484 |
ADDSUB %r9, %r11 C not on critical path
|
|
Packit |
5c3484 |
mov (vp), %r9 C not on critical path
|
|
Packit |
5c3484 |
setc R8(%rbx) C save carry out
|
|
Packit |
5c3484 |
mov %r12, %r13 C new for later
|
|
Packit |
5c3484 |
shl $63, %r12 C shift new right
|
|
Packit |
5c3484 |
shr %r14 C shift old left
|
|
Packit |
5c3484 |
mov %r10, (rp)
|
|
Packit |
5c3484 |
L(L11): ADDSUB %rax, %r11 C apply previous carry out
|
|
Packit |
5c3484 |
jc L(c0) C jump if ripple
|
|
Packit |
5c3484 |
L(rc0): mov 8(up), %r10
|
|
Packit |
5c3484 |
or %r14, %r12
|
|
Packit |
5c3484 |
ADDSUB %r9, %r8
|
|
Packit |
5c3484 |
mov 8(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rax)
|
|
Packit |
5c3484 |
mov %r11, %r14
|
|
Packit |
5c3484 |
shl $63, %r11
|
|
Packit |
5c3484 |
shr %r13
|
|
Packit |
5c3484 |
mov %r12, 8(rp)
|
|
Packit |
5c3484 |
L(L10): ADDSUB %rbx, %r8
|
|
Packit |
5c3484 |
jc L(c1)
|
|
Packit |
5c3484 |
L(rc1): mov 16(up), %r12
|
|
Packit |
5c3484 |
or %r13, %r11
|
|
Packit |
5c3484 |
ADDSUB %r9, %r10
|
|
Packit |
5c3484 |
mov 16(vp), %r9
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
mov %r8, %r13
|
|
Packit |
5c3484 |
shl $63, %r8
|
|
Packit |
5c3484 |
shr %r14
|
|
Packit |
5c3484 |
mov %r11, 16(rp)
|
|
Packit |
5c3484 |
L(L01): ADDSUB %rax, %r10
|
|
Packit |
5c3484 |
jc L(c2)
|
|
Packit |
5c3484 |
L(rc2): mov 24(up), %r11
|
|
Packit |
5c3484 |
or %r14, %r8
|
|
Packit |
5c3484 |
ADDSUB %r9, %r12
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
mov 24(vp), %r9
|
|
Packit |
5c3484 |
lea 32(vp), vp
|
|
Packit |
5c3484 |
setc R8(%rax)
|
|
Packit |
5c3484 |
mov %r10, %r14
|
|
Packit |
5c3484 |
shl $63, %r10
|
|
Packit |
5c3484 |
shr %r13
|
|
Packit |
5c3484 |
mov %r8, 24(rp)
|
|
Packit |
5c3484 |
lea 32(rp), rp
|
|
Packit |
5c3484 |
L(L00): ADDSUB %rbx, %r12
|
|
Packit |
5c3484 |
jc L(c3)
|
|
Packit |
5c3484 |
L(rc3): sub $4, n
|
|
Packit |
5c3484 |
ja L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): or %r13, %r10
|
|
Packit |
5c3484 |
ADDSUB %r9, %r11
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
mov %r12, %r13
|
|
Packit |
5c3484 |
shl $63, %r12
|
|
Packit |
5c3484 |
shr %r14
|
|
Packit |
5c3484 |
mov %r10, (rp)
|
|
Packit |
5c3484 |
L(cj3): ADDSUB %rax, %r11
|
|
Packit |
5c3484 |
jnc 1f
|
|
Packit |
5c3484 |
mov $1, R8(%rbx)
|
|
Packit |
5c3484 |
1: or %r14, %r12
|
|
Packit |
5c3484 |
mov %r11, %r14
|
|
Packit |
5c3484 |
shl $63, %r11
|
|
Packit |
5c3484 |
shr %r13
|
|
Packit |
5c3484 |
mov %r12, 8(rp)
|
|
Packit |
5c3484 |
L(cj2): or %r13, %r11
|
|
Packit |
5c3484 |
shl $63, %rbx
|
|
Packit |
5c3484 |
shr %r14
|
|
Packit |
5c3484 |
mov %r11, 16(rp)
|
|
Packit |
5c3484 |
L(cj1): or %r14, %rbx
|
|
Packit |
5c3484 |
mov %rbx, 24(rp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov R32(%r15), R32(%rax)
|
|
Packit |
5c3484 |
and $1, R32(%rax)
|
|
Packit |
5c3484 |
pop %r15
|
|
Packit |
5c3484 |
pop %r14
|
|
Packit |
5c3484 |
pop %r13
|
|
Packit |
5c3484 |
pop %r12
|
|
Packit |
5c3484 |
pop %rbx
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
L(c3): mov $1, R8(%rax)
|
|
Packit |
5c3484 |
jmp L(rc3)
|
|
Packit |
5c3484 |
EPILOGUE()
|