|
Packit |
5c3484 |
dnl AMD64 mpn_mul_basecase optimised for AMD Bulldozer and Piledriver.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Contributed to the GNU project by Torbjörn Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2003-2005, 2007, 2008, 2011-2013 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 mul_1 mul_2 mul_3 addmul_2
|
|
Packit |
5c3484 |
C AMD K8,K9
|
|
Packit |
5c3484 |
C AMD K10
|
|
Packit |
5c3484 |
C AMD bull ~4.8 ~4.55 - ~4.3
|
|
Packit |
5c3484 |
C AMD pile ~4.6 ~4.55 - ~4.55
|
|
Packit |
5c3484 |
C AMD bobcat
|
|
Packit |
5c3484 |
C AMD jaguar
|
|
Packit |
5c3484 |
C Intel P4
|
|
Packit |
5c3484 |
C Intel core
|
|
Packit |
5c3484 |
C Intel NHM
|
|
Packit |
5c3484 |
C Intel SBR
|
|
Packit |
5c3484 |
C Intel IBR
|
|
Packit |
5c3484 |
C Intel HWL
|
|
Packit |
5c3484 |
C Intel BWL
|
|
Packit |
5c3484 |
C Intel atom
|
|
Packit |
5c3484 |
C VIA nano
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C The inner loops of this code are the result of running a code generation and
|
|
Packit |
5c3484 |
C optimisation tool suite written by David Harvey and Torbjorn Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C TODO
|
|
Packit |
5c3484 |
C * Merge bull-specific mul_1, if it is not slower the TOOM22 range.
|
|
Packit |
5c3484 |
C Alternatively, we could tweak the present code (which was loopmixed for a
|
|
Packit |
5c3484 |
C different CPU).
|
|
Packit |
5c3484 |
C * Merge faster mul_2, such as the one in the same directory as this file.
|
|
Packit |
5c3484 |
C * Further micro-optimise.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C When playing with pointers, set this to $2 to fall back to conservative
|
|
Packit |
5c3484 |
C indexing in wind-down code.
|
|
Packit |
5c3484 |
define(`I',`$1')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`rp', `%rdi')
|
|
Packit |
5c3484 |
define(`up', `%rsi')
|
|
Packit |
5c3484 |
define(`un_param',`%rdx')
|
|
Packit |
5c3484 |
define(`vp', `%rcx')
|
|
Packit |
5c3484 |
define(`vn', `%r8')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`un', `%rbx')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`w0', `%r10')
|
|
Packit |
5c3484 |
define(`w1', `%r11')
|
|
Packit |
5c3484 |
define(`w2', `%r12')
|
|
Packit |
5c3484 |
define(`w3', `%r13')
|
|
Packit |
5c3484 |
define(`n', `%rbp')
|
|
Packit |
5c3484 |
define(`v0', `%r9')
|
|
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_basecase)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
IFDOS(` mov 56(%rsp), %r8d ')
|
|
Packit |
5c3484 |
push %rbx
|
|
Packit |
5c3484 |
push %rbp
|
|
Packit |
5c3484 |
mov un_param, un C free up rdx
|
|
Packit |
5c3484 |
neg un
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up), %rax C shared for mul_1 and mul_2
|
|
Packit |
5c3484 |
lea (up,un_param,8), up C point at operand end
|
|
Packit |
5c3484 |
lea (rp,un_param,8), rp C point at rp[un-1]
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (vp), v0 C shared for mul_1 and mul_2
|
|
Packit |
5c3484 |
mul v0 C shared for mul_1 and mul_2
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
test $1, R8(vn)
|
|
Packit |
5c3484 |
jz L(do_mul_2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(do_mul_1):
|
|
Packit |
5c3484 |
test $1, R8(un)
|
|
Packit |
5c3484 |
jnz L(m1x1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m1x0):mov %rax, w0 C un = 2, 4, 6, 8, ...
|
|
Packit |
5c3484 |
mov %rdx, w1
|
|
Packit |
5c3484 |
mov 8(up,un,8), %rax
|
|
Packit |
5c3484 |
test $2, R8(un)
|
|
Packit |
5c3484 |
jnz L(m110)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m100):lea 2(un), n C un = 4, 8, 12, ...
|
|
Packit |
5c3484 |
jmp L(m1l0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m110):lea (un), n C un = 2, 6, 10, ...
|
|
Packit |
5c3484 |
jmp L(m1l2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m1x1):mov %rax, w1 C un = 1, 3, 5, 7, ...
|
|
Packit |
5c3484 |
mov %rdx, w0
|
|
Packit |
5c3484 |
test $2, R8(un)
|
|
Packit |
5c3484 |
jz L(m111)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m101):lea 3(un), n C un = 1, 5, 9, ...
|
|
Packit |
5c3484 |
test n, n
|
|
Packit |
5c3484 |
js L(m1l1)
|
|
Packit |
5c3484 |
mov %rax, -8(rp)
|
|
Packit |
5c3484 |
mov %rdx, (rp)
|
|
Packit |
5c3484 |
pop %rbp
|
|
Packit |
5c3484 |
pop %rbx
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m111):lea 1(un), n C un = 3, 7, 11, ...
|
|
Packit |
5c3484 |
mov 8(up,un,8), %rax
|
|
Packit |
5c3484 |
jmp L(m1l3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(m1tp):mov %rdx, w0
|
|
Packit |
5c3484 |
add %rax, w1
|
|
Packit |
5c3484 |
L(m1l1):mov -16(up,n,8), %rax
|
|
Packit |
5c3484 |
adc $0, w0
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
add %rax, w0
|
|
Packit |
5c3484 |
mov w1, -24(rp,n,8)
|
|
Packit |
5c3484 |
mov -8(up,n,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, w1
|
|
Packit |
5c3484 |
adc $0, w1
|
|
Packit |
5c3484 |
L(m1l0):mul v0
|
|
Packit |
5c3484 |
mov w0, -16(rp,n,8)
|
|
Packit |
5c3484 |
add %rax, w1
|
|
Packit |
5c3484 |
mov %rdx, w0
|
|
Packit |
5c3484 |
mov (up,n,8), %rax
|
|
Packit |
5c3484 |
adc $0, w0
|
|
Packit |
5c3484 |
L(m1l3):mul v0
|
|
Packit |
5c3484 |
mov w1, -8(rp,n,8)
|
|
Packit |
5c3484 |
mov %rdx, w1
|
|
Packit |
5c3484 |
add %rax, w0
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
adc $0, w1
|
|
Packit |
5c3484 |
L(m1l2):mul v0
|
|
Packit |
5c3484 |
mov w0, (rp,n,8)
|
|
Packit |
5c3484 |
add $4, n
|
|
Packit |
5c3484 |
jnc L(m1tp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m1ed):add %rax, w1
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov w1, I(-8(rp),-24(rp,n,8))
|
|
Packit |
5c3484 |
mov %rdx, I((rp),-16(rp,n,8))
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dec R32(vn)
|
|
Packit |
5c3484 |
jz L(ret2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
lea 8(vp), vp
|
|
Packit |
5c3484 |
lea 8(rp), rp
|
|
Packit |
5c3484 |
push %r12
|
|
Packit |
5c3484 |
push %r13
|
|
Packit |
5c3484 |
push %r14
|
|
Packit |
5c3484 |
jmp L(do_addmul)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(do_mul_2):
|
|
Packit |
5c3484 |
define(`v1', `%r14')
|
|
Packit |
5c3484 |
push %r12
|
|
Packit |
5c3484 |
push %r13
|
|
Packit |
5c3484 |
push %r14
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov 8(vp), v1
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
test $1, R8(un)
|
|
Packit |
5c3484 |
jnz L(m2b1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m2b0):lea (un), n
|
|
Packit |
5c3484 |
mov %rax, w2 C 0
|
|
Packit |
5c3484 |
mov (up,un,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, w1 C 1
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
mov %rax, w0 C 1
|
|
Packit |
5c3484 |
mov w2, (rp,un,8) C 0
|
|
Packit |
5c3484 |
mov 8(up,un,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, w2 C 2
|
|
Packit |
5c3484 |
jmp L(m2l0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m2b1):lea 1(un), n
|
|
Packit |
5c3484 |
mov %rax, w0 C 1
|
|
Packit |
5c3484 |
mov %rdx, w3 C 2
|
|
Packit |
5c3484 |
mov (up,un,8), %rax
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
mov w0, (rp,un,8) C 1
|
|
Packit |
5c3484 |
mov %rdx, w0 C 3
|
|
Packit |
5c3484 |
mov %rax, w2 C 0
|
|
Packit |
5c3484 |
mov 8(up,un,8), %rax
|
|
Packit |
5c3484 |
jmp L(m2l1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(32)
|
|
Packit |
5c3484 |
L(m2tp):add %rax, w2 C 0
|
|
Packit |
5c3484 |
mov (up,n,8), %rax
|
|
Packit |
5c3484 |
adc $0, w0 C 1
|
|
Packit |
5c3484 |
L(m2l1):mul v0
|
|
Packit |
5c3484 |
add %rax, w2 C 0
|
|
Packit |
5c3484 |
mov (up,n,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, w1 C 1
|
|
Packit |
5c3484 |
adc $0, w1 C 1
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add w3, w2 C 0
|
|
Packit |
5c3484 |
adc $0, w1 C 1
|
|
Packit |
5c3484 |
add %rax, w0 C 1
|
|
Packit |
5c3484 |
mov w2, (rp,n,8) C 0
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, w2 C 2
|
|
Packit |
5c3484 |
adc $0, w2 C 2
|
|
Packit |
5c3484 |
L(m2l0):mul v0
|
|
Packit |
5c3484 |
add %rax, w0 C 1
|
|
Packit |
5c3484 |
mov %rdx, w3 C 2
|
|
Packit |
5c3484 |
adc $0, w3 C 2
|
|
Packit |
5c3484 |
add w1, w0 C 1
|
|
Packit |
5c3484 |
adc $0, w3 C 2
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add $2, n
|
|
Packit |
5c3484 |
mov w0, -8(rp,n,8) C 1
|
|
Packit |
5c3484 |
mov %rdx, w0 C 3
|
|
Packit |
5c3484 |
jnc L(m2tp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m2ed):add %rax, w2
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
add w3, w2
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov w2, I((rp),(rp,n,8))
|
|
Packit |
5c3484 |
mov %rdx, I(8(rp),8(rp,n,8))
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
add $-2, R32(vn)
|
|
Packit |
5c3484 |
jz L(ret5)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
lea 16(vp), vp
|
|
Packit |
5c3484 |
lea 16(rp), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(do_addmul):
|
|
Packit |
5c3484 |
push %r15
|
|
Packit |
5c3484 |
push vn C save vn in new stack slot
|
|
Packit |
5c3484 |
define(`vn', `(%rsp)')
|
|
Packit |
5c3484 |
define(`X0', `%r14')
|
|
Packit |
5c3484 |
define(`X1', `%r15')
|
|
Packit |
5c3484 |
define(`v1', `%r8')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(outer):
|
|
Packit |
5c3484 |
mov (vp), v0
|
|
Packit |
5c3484 |
mov 8(vp), v1
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up,un,8), %rax
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
test $1, R8(un)
|
|
Packit |
5c3484 |
jnz L(bx1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(bx0): mov %rax, X1
|
|
Packit |
5c3484 |
mov (up,un,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, X0
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
test $2, R8(un)
|
|
Packit |
5c3484 |
jnz L(b10)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b00): lea (un), n C un = 4, 8, 12, ...
|
|
Packit |
5c3484 |
mov (rp,un,8), w3
|
|
Packit |
5c3484 |
mov %rax, w0
|
|
Packit |
5c3484 |
mov 8(up,un,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, w1
|
|
Packit |
5c3484 |
jmp L(lo0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b10): lea 2(un), n C un = 2, 6, 10, ...
|
|
Packit |
5c3484 |
mov (rp,un,8), w1
|
|
Packit |
5c3484 |
mov %rdx, w3
|
|
Packit |
5c3484 |
mov %rax, w2
|
|
Packit |
5c3484 |
mov 8(up,un,8), %rax
|
|
Packit |
5c3484 |
jmp L(lo2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(bx1): mov %rax, X0
|
|
Packit |
5c3484 |
mov (up,un,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, X1
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
test $2, R8(un)
|
|
Packit |
5c3484 |
jz L(b11)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b01): lea 1(un), n C un = 1, 5, 9, ...
|
|
Packit |
5c3484 |
mov (rp,un,8), w2
|
|
Packit |
5c3484 |
mov %rdx, w0
|
|
Packit |
5c3484 |
mov %rax, w3
|
|
Packit |
5c3484 |
jmp L(lo1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b11): lea -1(un), n C un = 3, 7, 11, ...
|
|
Packit |
5c3484 |
mov (rp,un,8), w0
|
|
Packit |
5c3484 |
mov %rax, w1
|
|
Packit |
5c3484 |
mov 8(up,un,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, w2
|
|
Packit |
5c3484 |
jmp L(lo3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(32)
|
|
Packit |
5c3484 |
L(top):
|
|
Packit |
5c3484 |
L(lo2): mul v0
|
|
Packit |
5c3484 |
add w1, X1
|
|
Packit |
5c3484 |
mov X1, -16(rp,n,8)
|
|
Packit |
5c3484 |
mov %rdx, X1
|
|
Packit |
5c3484 |
adc %rax, X0
|
|
Packit |
5c3484 |
adc $0, X1
|
|
Packit |
5c3484 |
mov -8(up,n,8), %rax
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
mov -8(rp,n,8), w1
|
|
Packit |
5c3484 |
mov %rdx, w0
|
|
Packit |
5c3484 |
add w1, w2
|
|
Packit |
5c3484 |
adc %rax, w3
|
|
Packit |
5c3484 |
adc $0, w0
|
|
Packit |
5c3484 |
L(lo1): mov (up,n,8), %rax
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
add w2, X0
|
|
Packit |
5c3484 |
mov X0, -8(rp,n,8)
|
|
Packit |
5c3484 |
mov %rdx, X0
|
|
Packit |
5c3484 |
adc %rax, X1
|
|
Packit |
5c3484 |
mov (up,n,8), %rax
|
|
Packit |
5c3484 |
adc $0, X0
|
|
Packit |
5c3484 |
mov (rp,n,8), w2
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add w2, w3
|
|
Packit |
5c3484 |
adc %rax, w0
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, w1
|
|
Packit |
5c3484 |
adc $0, w1
|
|
Packit |
5c3484 |
L(lo0): mul v0
|
|
Packit |
5c3484 |
add w3, X1
|
|
Packit |
5c3484 |
mov X1, (rp,n,8)
|
|
Packit |
5c3484 |
adc %rax, X0
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, X1
|
|
Packit |
5c3484 |
adc $0, X1
|
|
Packit |
5c3484 |
mov 8(rp,n,8), w3
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add w3, w0
|
|
Packit |
5c3484 |
adc %rax, w1
|
|
Packit |
5c3484 |
mov 16(up,n,8), %rax
|
|
Packit |
5c3484 |
mov %rdx, w2
|
|
Packit |
5c3484 |
adc $0, w2
|
|
Packit |
5c3484 |
L(lo3): mul v0
|
|
Packit |
5c3484 |
add w0, X0
|
|
Packit |
5c3484 |
mov X0, 8(rp,n,8)
|
|
Packit |
5c3484 |
mov %rdx, X0
|
|
Packit |
5c3484 |
adc %rax, X1
|
|
Packit |
5c3484 |
adc $0, X0
|
|
Packit |
5c3484 |
mov 16(up,n,8), %rax
|
|
Packit |
5c3484 |
mov 16(rp,n,8), w0
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
mov %rdx, w3
|
|
Packit |
5c3484 |
add w0, w1
|
|
Packit |
5c3484 |
adc %rax, w2
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
mov 24(up,n,8), %rax
|
|
Packit |
5c3484 |
add $4, n
|
|
Packit |
5c3484 |
jnc L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): mul v0
|
|
Packit |
5c3484 |
add w1, X1
|
|
Packit |
5c3484 |
mov X1, I(-16(rp),-16(rp,n,8))
|
|
Packit |
5c3484 |
mov %rdx, X1
|
|
Packit |
5c3484 |
adc %rax, X0
|
|
Packit |
5c3484 |
adc $0, X1
|
|
Packit |
5c3484 |
mov I(-8(up),-8(up,n,8)), %rax
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
mov I(-8(rp),-8(rp,n,8)), w1
|
|
Packit |
5c3484 |
add w1, w2
|
|
Packit |
5c3484 |
adc %rax, w3
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
add w2, X0
|
|
Packit |
5c3484 |
adc $0, X1
|
|
Packit |
5c3484 |
mov X0, I(-8(rp),-8(rp,n,8))
|
|
Packit |
5c3484 |
add w3, X1
|
|
Packit |
5c3484 |
mov X1, I((rp),(rp,n,8))
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov %rdx, I(8(rp),8(rp,n,8))
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
addl $-2, vn
|
|
Packit |
5c3484 |
lea 16(vp), vp
|
|
Packit |
5c3484 |
lea 16(rp), rp
|
|
Packit |
5c3484 |
jnz L(outer)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
pop %rax C deallocate vn slot
|
|
Packit |
5c3484 |
pop %r15
|
|
Packit |
5c3484 |
L(ret5):pop %r14
|
|
Packit |
5c3484 |
pop %r13
|
|
Packit |
5c3484 |
pop %r12
|
|
Packit |
5c3484 |
L(ret2):pop %rbp
|
|
Packit |
5c3484 |
pop %rbx
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|
|
rpm-build |
01f633 |
CF_PROT
|